There are three broad ways to use it. You can run the weights yourself, which can fit on one high-memory GPU if you keep the context short. You can call Google's Gemini API or another host. Or you can use it in Maple, where we offer it on Pro, Max and Team plans with the full 262,144-token window, through our protected inference path.

This guide separates what belongs to the model from what belongs to each route, so you can pick one.

What is Gemma 4 31B?

It is the dense member of a family that now has five sizes. Dense means every parameter works on every token, so the whole model has to sit in memory and all of it runs each step. That makes 31B slower than its mixture-of-experts sibling, 26B A4B, and generally stronger in Google's own results.

Its attention design is what makes the long window affordable. Of its 60 layers, 50 look back only over the last 1,024 tokens, and every sixth layer reads the whole context. Google describes this as interleaving "local sliding window attention with full global attention". Once a conversation grows long, most of its memory therefore sits in just 10 layers, which matters when you size a machine.

Images go through a vision encoder of about 550 million parameters. Each image costs 70 to 1,120 tokens depending on the resolution budget, 280 by default. Video is read as frames, and the 31B model has no audio input. Google lists more than 35 languages as supported out of the box, from pre-training on more than 140.

Is Gemma 4 31B open source?

The weights are. Google publishes them under the Apache 2.0 license, a standard permissive license, and the repository needs no click-through. Earlier Gemma releases used Google's own Gemma Terms of Use; for Gemma 4 those terms simply point to Apache 2.0.

Google also published a technical report.⁵ It does not publish its training data, which has a January 2025 cutoff. If your definition of open source requires the data, Gemma 4 31B is open-weight rather than fully open.

Which Gemma 4 model is 31B, and what changed since launch?

Google released Gemma 4 in four sizes (E2B, E4B, 26B A4B and 31B) and added a fifth, the 12B Unified model, two months later.⁶ Pages written at launch describe four sizes and say audio belongs to the two small models. Both are now out of date: audio is supported on E2B, E4B and 12B, and never on 31B.

The 31B weights themselves have not changed since launch. What Google has added around them are tools for running them more cheaply. Its 4-bit quantization-aware training (QAT) checkpoints were trained to hold up at low precision, so they are a better starting point than a generic 4-bit conversion. Its MTP drafter models speed up decoding with speculative decoding; Google says by "up to 3x" with the same output quality.

Gemma 4 31B or 26B A4B?

The 26B A4B is a mixture-of-experts model: 25.2 billion parameters in total, 3.8 billion active per token. It has the same window and input types as 31B. Its memory saving is smaller than its name suggests, because every expert still has to be loaded.

The benchmark rows are Google's results for the instruction-tuned models; Google chose the tests and settings, and these are not Maple measurements. The gap is narrow on knowledge and reasoning and wide on long-context retrieval. If your work depends on finding details deep in a long document, 31B is the safer choice. If you want faster answers on modest hardware, 26B A4B is worth testing on your own tasks.

How much memory does Gemma 4 31B need?

Two amounts, and most hardware advice merges them. The first is the weights, which Google publishes. The second is the working memory for the conversation, the KV cache, which grows with every token in the context.

Google is explicit that its table covers only the first:

The estimates in the preceding table only account for the memory required to load the static model weights. They don't include the additional VRAM needed for supporting software or the context window.

We estimated the second from the model's published configuration.⁷ The 50 sliding-window layers need a fixed 0.84 GB at 16-bit precision, whatever the context length. The 10 global layers need about 41 KB per token, because Gemma stores one shared value for keys and values there. If a serving engine stores keys and values separately, that doubles to about 82 KB.

The totals add Google's 17.5 GB load figure for 4-bit weights; for 8-bit, add the KV column to 34.9 GB. They are for one conversation at 16-bit cache precision, on an engine that keeps the sliding-window layers to their window. Add the engine's own overhead. An 8-bit cache halves the KV column, and an engine that allocates the full window to every layer needs far more.

Read the table this way. A 24 GB GPU runs 4-bit Gemma 4 31B comfortably at 32K tokens, but not at the full 256K. Google also notes that Ollama defaults the context to 4K on machines with less than 24 GiB of GPU memory.⁸ A 48 GB card holds 8-bit weights with up to about 128K tokens, and the BF16 download of 62.55 GB needs an 80 GB-class GPU.⁹

How do you run Gemma 4 31B yourself?

On one GPU, in most setups. Running the weights yourself gives you the most control over the model, the serving stack and where your data goes. The cost is hardware, setup and keeping the stack current.

The common routes, from easiest to most configurable:

• Ollama or LM Studio. The gemma4:31b tag in Ollama is 19 to 20 GB, and both apps list a QAT build. Set the context length yourself, since Ollama's default is short on smaller GPUs.
• llama.cpp. Google's guide recommends its QAT Q4_0 GGUF, an 18.85 GB download including the 1.20 GB vision projector. It warns that images consume context quickly.
• vLLM. The vLLM project's recipe lists one 80 GB NVIDIA GPU as the minimum for BF16 and gives a two-GPU example capped at a 32K window. It needs the gemma4 reasoning and tool-call parsers for thinking and function calling, and Google's 4-bit w4a16 checkpoint cuts the weights from 59.0 to 19.8 GB.¹⁰

Self-hosting is the right call when:

• your information must never leave hardware you control
• you need to fine-tune or deeply instrument the model
• you already have a GPU with 24 GB or more and only need modest context

If you need the full window, or you work across several devices, a hosted route avoids buying for the peak. Our DeepSeek V4.1 Flash guide covers a model where that choice is starker, because its checkpoint is 510 GB.

How do you use Gemma 4 31B in Maple?

Choose Gemma 4 31B under More models in the model picker in Maple Research, on a Pro, Max or Team plan. It is not available on Free. We configure it with the full 262,144-token window, so a long report, a codebase or a stack of notes can go into one conversation without a GPU of your own.

In the app, that window holds both your conversation and the answer. When a long chat outgrows it, Maple trims older turns from the middle first and always keeps your first message and your newest one. Thinking is on by default, and you can switch it off for quicker replies.

You can upload an image while using Gemma 4 31B in the app. The app first has a vision model describe the image, then gives that description to the model you selected.¹¹ Gemma 4 31B answers from the description, not a direct view of the image.

For developer tools and existing applications, Maple Proxy is our API route, with the model ID gemma4-31b. There, an image you send goes to Gemma 4 31B itself, so use Proxy when a task depends on a small visual detail. Thinking is off on the API unless you set reasoning_effort, and the Proxy model list shows the current window and input types.

Plan details are on our pricing page, and Gemma 4 31B sits beside the other models we offer in our model library. If you need a longer window, our GLM-5.3 Flash guide covers a model with a million tokens.

The route has limits to plan around:

• Images are described in the app. Use Maple Proxy when the model must see the image itself.
• A paid plan is required. Gemma 4 31B is on Pro, Max and Team.
• Long chats are trimmed. Past 262,144 tokens, the middle of a conversation goes first.

What information reaches each route?

That depends on who operates the model: you, a conventional host, or Maple. Three questions get blurred together. Who can read your prompts while they are processed, what happens to them afterward, and is any of it used for training?

Self-hosted weights. Your prompts and outputs stay on infrastructure you operate. The trust boundary is your own access controls, logging and retention. That makes this the strongest option when information must stay on your hardware, and it makes you responsible for securing it.

Google's Gemini API. Google hosts Gemma 4 31B as gemma-4-31b-it. Its pricing page lists Gemma 4 as free of charge, with no paid tier.¹² Google's terms for unpaid services say it uses what you submit to improve its products and machine learning technologies, and that "human reviewers may read, annotate, and process your API input and output."¹³

The same terms tell developers directly: "Do not submit sensitive, confidential, or personal information to the Unpaid Services." It is a clear disclosure. Users in the European Economic Area, Switzerland and the UK get Google's paid terms, under which prompts are not used to improve its products. Other hosts that serve the open weights set their own terms, so read them for the service you would use.

Maple. We run Google's open weights inside an attested enclave rather than calling Google's API. Your requests are encrypted before they leave your device, and our proof page explains the attestation you can check yourself.

The model runs in its own attested enclave, and Maple's enclave verifies it before forwarding a prompt. It works like certificate chains on the web: you check Maple's enclave, and Maple's enclave checks the model's. Our privacy notice covers everything outside your chats.

That design has a scope. It protects the contents of your chats: your prompts and the responses. It doesn't make information private if you've already shared it elsewhere.

It is also separate from the safety of what the model's output leads you to do. If you paste its code into production or act on its advice, confidential processing doesn't vouch for that code or those actions. Review outputs as you would any untrusted input.

Which route fits your task?

It depends on where your information may go, how much context you need and what hardware you have. The table compares the three routes on the same terms.

A rough decision rule:

• Choose self-hosting if information must stay on your hardware and your context needs are modest.
• Choose the Gemini API for prompts you would be comfortable with Google using and reviewing.
• Choose Maple if you want the full 262K window on sensitive material without buying a GPU for it.

Sources

1. Google DeepMind, Gemma 4 31B IT model card, last updated July 15, 2026. Source for the parameter count, layers, attention design, context length, input types, vision encoder, image token budgets, video limit, languages, training cutoff, thinking, function calling, benchmark results and the 26B A4B specifications.
2. Google, Gemma 4 license (Apache License 2.0), April 1, 2026, and the Gemma Terms of Use, which point Gemma 4 to it.
3. Google, Gemma 4 model overview. Source for the memory table for every Gemma 4 size and its scope, the 26B A4B memory figures and the quotation above.
4. Maple, public model configuration and app model picker, with context handling and the thinking default. Source for the plans, configured window, trimming and thinking settings in Maple.
5. Gemma Team, Google DeepMind, “Gemma 4 Technical Report”, arXiv, submitted July 2, 2026.
6. Google, Gemma releases, also published on ai.google.dev, updated October 7, 2026. Source for the release dates of Gemma 4, the MTP drafters and the 12B Unified model; Google's QAT announcement (June 5, 2026) dates the 4-bit checkpoints, and Google's MTP drafter card for 31B is the source for the speedup.
7. Maple's calculation from Google's Gemma 4 31B configuration: 10 global layers with 4 key-value heads of 512 dimensions and shared keys and values, and 50 sliding-window layers of 1,024 tokens with 16 key-value heads of 256 dimensions, at 2 bytes per value.
8. Google, Run Gemma with Ollama. Source for Ollama's input types and its 4K default context under 24 GiB of GPU memory; the Ollama library lists the tag sizes.
9. Hugging Face, repository file listings for Gemma 4 31B IT and Google's QAT Q4_0 GGUF. Our sums of the weight files: 62.55 GB at BF16 and 18.85 GB for the GGUF with its vision projector.
10. vLLM project, Gemma 4 serving recipe, updated October 17, 2026. Source for the minimum GPU, the two-GPU example, the required parsers and the QAT memory saving.
11. Maple, image description in the app and native images on the API. Source for how each surface handles an uploaded image.
12. Google, Gemini Developer API pricing, updated October 19, 2026. Source for Gemma 4's free tier and the absence of a paid tier.
13. Google, Gemini API Additional Terms of Service, updated April 28, 2026. Source for how Google uses and reviews content submitted to unpaid services, the quoted instruction, and the paid terms for the EEA, Switzerland and the UK.

Frequently Asked Questions

Is Gemma 4 31B open source?

The weights are released under the Apache 2.0 license, with no click-through and no use policy attached, so you can download, modify and build on them commercially. Google also publishes a technical report. It does not publish its training data.

Can Gemma 4 31B run on a 24 GB GPU?

With 4-bit weights and a modest context, yes. Google estimates 17.5 GB to load the model at Q4_0, and by our estimate a 32K-token conversation adds about 2 to 3.5 GB. The full 256K window adds 12 to 22 GB, which does not fit beside the weights on a 24 GB card.

Does Gemma 4 31B accept audio or video?

No audio. The 31B model takes text and images and returns text. It can read video only as a series of frames, up to 60 seconds at one frame per second. Google's smaller E2B, E4B and 12B models accept audio.

What context length does Maple give Gemma 4 31B?

The full 262,144-token window, in the Maple app and on Maple Proxy. In the app the window holds both your conversation and the answer, and when a long chat outgrows it, older turns from the middle are trimmed first.

Can Gemma 4 31B see my images in Maple?

Through Maple Proxy, yes: an image you send goes to Gemma 4 31B itself. In the Maple app, a vision model first describes each uploaded image, and the model you picked works from that description.

Can Maple read my Gemma 4 31B chats?

No. No one at Maple can read your chats. They are encrypted on your device and decrypted only inside a hardware-isolated enclave running open-source code you can verify, and your chat history is stored encrypted. Our privacy notice covers everything outside your chats.

OpenAI publishes the weights but doesn't run the model for you: GPT-OSS is not in ChatGPT or OpenAI's API. That leaves three routes. You can run it on a GPU with at least 60 GB of memory, use a service that hosts it, or use it in Maple, where it is on the Free plan with the full 131,072-token window, through our protected inference path.

This guide separates what belongs to the model from what belongs to each route, so you can pick one.

What is GPT-OSS 120B?

It is a sparse model. Each of its 36 layers holds 128 experts, and a router picks four of them for every token, which is why only 5.1 billion of its 116.8 billion parameters are active for each token.

It does not shrink the memory you need. Every expert has to be available, because the next token may route to any of them. OpenAI's answer is MXFP4: it post-trained the model with the expert weights in this 4-bit format (4.25 bits per parameter), which brings the whole checkpoint to 60.8 GiB, a download of about 65 GB.

The model works through a problem before answering, and you can set how hard it thinks: low, medium or high reasoning. It was trained to use tools such as web browsing, a Python environment and functions a developer defines, but only when the app running it provides them. Its knowledge cutoff is June 2024, so it doesn't know about later events unless a tool supplies them.

OpenAI also notes that the model's raw reasoning isn't meant for end users. The card says it can contain hallucinated content and language that doesn't reflect OpenAI's safety policies, so developers should filter or summarize it before showing it.

Is GPT-OSS 120B open source?

The weights are. OpenAI publishes them on Hugging Face under the Apache 2.0 license, a standard permissive license, with reference code for running them. The accompanying usage policy is two sentences:

We aim for our tools to be used safely, responsibly, and democratically, while maximizing your control over how you use them. By using OpenAI gpt-oss-120b, you agree to comply with all applicable law.

OpenAI itself calls the models open-weight: the trained weights are public, while the model card describes the training data only as "a text-only dataset with trillions of tokens." If your definition of open source includes the data, GPT-OSS 120B doesn't meet it.

Can you use GPT-OSS 120B in ChatGPT or OpenAI's API?

No. OpenAI's Help Center is direct about it, for both GPT-OSS models and the Safeguard variants:²

These models are not served through the OpenAI API and are not available in ChatGPT.

OpenAI's developer docs also have a model page for GPT-OSS 120B listing its specifications. Use it for the specs. For access, you run the weights yourself or use a service that hosts them.

How good is GPT-OSS 120B?

In OpenAI's own evaluations, close to its small proprietary reasoning models. The model card's summary is that it "surpasses OpenAI o3-mini and approaches OpenAI o4-mini accuracy." These are OpenAI's results: it chose the benchmarks and ran them at the high reasoning setting, and they are not Maple measurements.

The pattern is consistent: slightly behind o4-mini on most tests, ahead of o3-mini on all five, and ahead of both on HealthBench. OpenAI adds that the models do not replace a medical professional and are not intended for diagnosing or treating disease.

The card is also candid about weaknesses. On SimpleQA, a test of short factual questions, it answers correctly less often than o4-mini and hallucinates more. Check its factual claims, especially about anything after June 2024.

Should you use GPT-OSS 120B or 20B?

OpenAI released two sizes together, and the choice comes down to your hardware. OpenAI says the 20B can run on systems with as little as 16 GB of memory; the 120B is the stronger model.

If your machine has less than 60 GB of GPU or unified memory, 20B is the realistic choice. If you can reach that much memory or use a hosted route, 120B gives you the stronger model.

What hardware does GPT-OSS 120B need to self-host?

At least 60 GB of GPU memory, or of unified memory on a workstation, in OpenAI's own guides. Running the weights yourself gives you the most control over the model, the serving stack and where your data goes. The cost is hardware and setup.

We checked the download size ourselves. Hugging Face's file listing for the repository shows 15 weight files totaling 65.25 GB, which is 60.8 GiB in the binary units the model card uses.³ The repository also holds two more versions of the weights for OpenAI's reference implementations, in original/ and metal/, for 196 GB in all; most runtimes need only the first set.

OpenAI's Transformers recipe says the 120B model "requires ≥60GB VRAM or multi-GPU setup" and calls it ideal for H100-class hardware.⁴ In Transformers, its MXFP4 weights need NVIDIA's Hopper generation or later, which includes the H100 and the RTX 50 series. OpenAI's Ollama guide, written for PCs and Macs, says the 120B is "best with ≥60GB VRAM or unified memory."⁷

The vLLM recipe, an open-source serving project's guide, says the model fits on a single A100 80GB.⁸ It also lists the hardware it has verified, from NVIDIA H100 and B200 to AMD MI300X and a set of four Intel Arc Pro B70 cards.

NVIDIA's GeForce RTX 5090 carries 32 GB and the RTX 4090 24 GB.⁹ Neither holds GPT-OSS 120B on its own. OpenAI's Ollama guide says you can offload part of the model to the CPU when VRAM runs short, but expect it to run slower.

You may see community versions in other formats. Check their file size before assuming they help: OpenAI's expert weights are already 4.25 bits each, and its Ollama guide notes that the models ship in MXFP4 with no other quantization from OpenAI.

Self-hosting is the right call when:

• your information must never leave hardware you control
• you need to fine-tune, modify or deeply instrument the model
• you already run GPU servers and have capacity to spare

If none of those apply, ask whether the operational work buys you anything a hosted route can't. Our DeepSeek V4.1 Flash guide works through the same question for a far larger open-weight model.

How do you use GPT-OSS 120B in Maple?

Sign up and choose GPT-OSS 120B from the model picker in Maple Research. On the Free plan it is also the model behind Quick, the picker's option for fast, everyday answers. We configure it with the full 131,072-token window, the same as the upstream model, so a long document or a long working session fits in one conversation.

The composer's thinking control sets its reasoning level to low, medium or high; medium is the default. Use high for multi-step math, code or analysis, and low when you want a fast answer.

Plan details are on our pricing page. GPT-OSS 120B appears in our model library beside the other models we offer. On paid plans, Quick moves to GLM-5.3 Flash, which our GLM-5.3 Flash guide covers, and GPT-OSS 120B stays in the picker.

For developer tools and existing applications, Maple Proxy is our API route, using the model ID gpt-oss-120b. API access comes with Pro, Max and Team. Check the Proxy model list for its current context limit and features before you build on it.

This route gives you two things. You get a model that needs 60 GB of GPU memory without buying or renting that GPU. And your requests run through our protected inference path rather than a conventional hosted API.

The route has limits to plan around:

• Text only. GPT-OSS 120B doesn't read images. On paid plans, Research has another model describe an uploaded image first, and GPT-OSS 120B answers from that description. The Free plan doesn't include image upload.
• Knowledge stops in June 2024. For anything newer, turn on web search in the composer. On Free, it can run up to five searches per answer.
• The API needs a paid plan. The model is free in the app; API access starts at Pro.

What information reaches each route?

That depends on who operates the model: you, a hosting service, or Maple. Three questions get blurred together here. Who can read your prompts while they are processed, what happens to them afterward, and is any of it used for training?

Self-hosted weights. Your prompts and outputs stay on infrastructure you operate. OpenAI's Help Center notes that OpenAI doesn't receive the data you send to a copy you host yourself, unless you choose to share it. The trust boundary is your own: your access controls, logging and retention, which makes this the strongest option when information must stay on your hardware.

A hosting service. OpenAI partnered with deployment platforms ahead of launch, and many services now run GPT-OSS 120B.¹⁰ Each sets its own terms. Read them for the service and plan you would use: whether prompts are retained, who can access them, and whether they are used for training.

Maple. We run OpenAI's open weights inside an attested enclave. Your requests are encrypted before they leave your device, and our proof page explains the attestation you can check yourself.

The model runs in its own attested enclave, and Maple's enclave verifies it before forwarding a prompt. It works like certificate chains on the web: you check Maple's enclave, and Maple's enclave checks the model's. Our privacy notice covers everything outside your chats.

That design has a scope. It protects the contents of your chats: your prompts and the responses. It doesn't make information private if you've already shared it elsewhere.

It is also separate from the quality of the answer. A private route doesn't make the model's output more accurate, so review its code, facts and suggestions as you would any untrusted input.

How is GPT-OSS Safeguard different?

GPT-OSS Safeguard 120B is a separate model with a narrower job. OpenAI released it on October 29, 2025, as a research preview fine-tuned from GPT-OSS, under the same Apache 2.0 license.¹¹ You give it a written policy and the content to check, and it returns a verdict with its reasoning.

It is built for moderation and trust-and-safety pipelines, not conversation; OpenAI recommends against making it the core of what end users interact with. We offer it as gpt-oss-safeguard-120b through Maple Proxy only, not in the Research picker.

Which route fits your task?

It depends on where your information may go, how much hardware you can run and whether you need the API. The table compares the three routes on the same terms.

A rough decision rule:

• Choose self-hosting if information must stay on your hardware, you need to fine-tune the weights, or you already have 60 GB of memory to spare.
• Choose a hosting service if your prompts aren't sensitive and its terms fit your data.
• Choose Maple if you want GPT-OSS 120B without a 60 GB GPU and without a host that can read your chats.

Sources

1. OpenAI, gpt-oss-120b & gpt-oss-20b Model Card, August 5, 2025. Source for the architecture, parameter and expert counts, MXFP4, checkpoint size, context length, knowledge cutoff, training data description, reasoning levels, tools, the chain-of-thought guidance, the benchmark and SimpleQA results at high reasoning, and the medical caveat.
2. OpenAI Help Center, “OpenAI open-weight models (gpt-oss)”. Source for the models not being served in ChatGPT or the OpenAI API, the open-weight terminology, free use under Apache 2.0 and the usage policy, and self-hosted data not reaching OpenAI.
3. Hugging Face, repository file listing for openai/gpt-oss-120b. Our sum of the 15 safetensors files (65.25 GB, 60.8 GiB) and of the repository's original/ and metal/ weight files (196 GB in all).
4. OpenAI, “How to run gpt-oss with Transformers”, OpenAI Cookbook, August 5, 2025. Source for the 60 GB memory requirement, H100-class hardware and MXFP4 support on Hopper or later.
5. Maple, public model configuration and Research model picker. Source for the Free plan tier, Quick on Free, the configured context window, text-only input and the reasoning levels in Maple.
6. OpenAI, gpt-oss-20b model card on Hugging Face and repository file listing. Source for the 20B's parameter counts and our sum of its weight files (13.8 GB).
7. OpenAI, “How to run gpt-oss locally with Ollama”, OpenAI Cookbook, August 5, 2025. Source for 60 GB of VRAM or unified memory, CPU offload and MXFP4 as the only quantization OpenAI ships.
8. vLLM project, GPT-OSS serving recipe (source file), updated September 22, 2026. Source for the single A100 80GB fit and the verified hardware, including four Intel Arc Pro B70 cards.
9. NVIDIA, GeForce RTX 5090, GeForce RTX 4090 and A100 product pages. Source for 32 GB, 24 GB and 80 GB of GPU memory.
10. OpenAI, “Introducing gpt-oss”, August 5, 2025. Source for the release date and the deployment partners at launch.
11. OpenAI, “Introducing gpt-oss-safeguard”, October 29, 2025, and the gpt-oss-safeguard technical report. Source for the release, license, inputs and outputs, and OpenAI's recommended use.

Frequently Asked Questions

Is GPT-OSS 120B free?

The weights are free to download and use under the Apache 2.0 license and OpenAI's short gpt-oss usage policy, which asks you to comply with applicable law. Running them is not free: you pay for the GPU, or for a hosted service. In Maple, GPT-OSS 120B is available on the Free plan.

How much VRAM does GPT-OSS 120B need?

OpenAI's Transformers recipe asks for at least 60 GB of GPU memory, or several GPUs, with the model's native MXFP4 weights; its Ollama guide says at least 60 GB of VRAM or unified memory. In practice that means one 80 GB data-center GPU such as an H100 or A100 80GB, several GPUs, or a workstation with that much unified memory.

Is GPT-OSS 120B or 20B better?

120B is the stronger model in OpenAI's evaluations, and 20B is the one that fits ordinary hardware: OpenAI says it runs in about 16 GB of memory. If you have 60 GB or more of GPU or unified memory, or use a hosted route, choose 120B. If your machine has less, choose 20B.

Can I run GPT-OSS 120B locally with Ollama?

Yes, if your machine has enough memory. OpenAI's Ollama guide says the 120B is best with at least 60 GB of VRAM or unified memory, and that you can offload to the CPU if you're short on VRAM, though it runs slower. The download is about 65 GB whichever runtime you use.

Is GPT-OSS 120B in ChatGPT or the OpenAI API?

No. OpenAI's Help Center says the gpt-oss models are not served through the OpenAI API and are not available in ChatGPT. To use GPT-OSS 120B you run the weights yourself or use a service that hosts them, such as Maple.

Can Maple read my GPT-OSS 120B chats?

No. No one at Maple can read your chats. They are encrypted on your device and decrypted only inside a hardware-isolated enclave running open-source code you can verify, and your chat history is stored encrypted. Our privacy notice covers everything outside your chats.

We make Maple. Here's exactly how we judged every app, so you can check each claim.

Private chats are one part of using AI privately. Our guide to using AI privately covers your connection, your email and how you pay.

What a private AI chat app is (and isn't)

A private AI chat app limits who can read what you type. Every hosted app here encrypts your connection; they differ in what happens once your prompt arrives, and in what evidence you get.

Apps sold as private fall into five privacy models:

1. Policy only. The company can read your prompts, and its policy limits what it does with them. ChatGPT and Claude work this way.
2. Anonymizing proxy. A middleman strips your identity and forwards readable text to a model provider. Duck.ai and most of Brave Leo work this way.
3. Operator no-logs. The operator runs the model, sees your prompts in use and promises not to keep them. Proton Lumo and Venice's default mode work this way.
4. Attested enclave. Prompts are decrypted only inside hardware-isolated enclaves, and your app checks the enclave's code by remote attestation before sending. Maple, Confer and Privatemode work this way.
5. Local. The model runs on your own hardware, so there is no company in the loop.

Only the attested enclave model does both for a hosted app. The first three ask you to trust a promise; the fourth lets you check. Private also doesn't mean untraceable: an app that can't read your chats can still know which account sent them and how you paid.

Want to see the difference? Try Maple free, and check the enclave before you send anything.

How we compared them

We read each company's own security pages, policies, pricing and app listings, and every entry links them. We didn't run lab tests of the other apps.

The first two criteria decide whether an app makes the list. The other four are compared for every app that passes, and the list is ordered by how many platforms each one supports.

1. Can the company read your chats?

   By design, not by policy.

2. Can you verify it yourself?

   Remote attestation and published code.

3. Does it train on your data by default?

   Read from each company's own policy.

4. Model choice

   How many models, and which kind.

5. Price and free tier

   The entry paid plan, and what's free.

6. Platforms

   Web, phone, desktop and API.

The private AI chat apps that pass both tests

Maple AI: private chat on every device, plus an API

• Reads your chats: No
• Verify yourself: Yes
• Trains by default: No
• Models: 7 open-weight
• Price: Free; Pro $20/mo
• Platforms: Web, iOS, Android, macOS, Windows, Linux, API

Maple cannot see your chats. They are encrypted on your device and decrypted only inside a hardware-isolated enclave running open-source code you can verify. The model runs in its own attested enclave: you check Maple's enclave, and Maple's enclave checks the model's before it forwards a prompt, much as certificate chains work on the web.

Before the app sends anything, it checks the enclave's attestation; if a check fails, nothing is sent. The code is on GitHub under the MIT license. You still rely on the hardware behind each enclave, the published code and Maple's signing key.

It passes both tests, and it's the only app that does so with apps for every major platform, an OpenAI-compatible API and a free plan. Maple offers seven open-weight chat models, including GLM 5.3, Kimi K3 and DeepSeek V4.1 Flash. You can sign up without an email and pay with Bitcoin, as our guide to using AI privately explains.

Where Maple falls short: open-weight models only, so no GPT or Claude, and no BAA or independent audit yet; we're working toward a BAA and a SOC 2 audit.

Best suited for: anyone who wants private AI chat on every device, with a privacy claim they can check.

Checked 15 Oct 2026.

Confer: an enclave chat app for web and iPhone

• Reads your chats: No
• Verify yourself: Yes
• Trains by default: No
• Models: Open models, not named; no model picker
• Price: 20 messages/day free; $34.99/mo
• Platforms: Web, iOS (also runs on Apple silicon Macs)

Confer encrypts prompts from your device into a trusted execution environment, after the client checks its attestation against a logged release. Its builds are bit-for-bit reproducible, and it publishes the source for its server image and proxy so anyone can rebuild and compare, per Confer's private inference post. In its words, "Confer cannot read your prompts or responses—not by policy, but by design."⁵

Its privacy policy says history is stored encrypted and not used for training. Confer was created by Signal founder Moxie Marlinspike, who wrote in March 2026 that he will work to integrate Confer's privacy technology into Meta AI, while Confer continues as an independent company; no timeline was given.

Sign-up is open. Confer launched on the web in December 2025 and on iOS in May 2026, and the free Guest plan includes 20 messages a day and 5 saved chats.

Where it falls short: Membership costs $34.99 a month, the most expensive plan here. Confer uses open models and doesn't list which ones, and there's no model picker. There's no public API, and an Android app is "in the works." Parts of its server code are published on GitHub,⁹ but its apps aren't open source.

Best suited for: people who want an enclave app from Signal's founder and don't need to choose a model.

Checked 15 Oct 2026.

Privatemode: attested AI with an API, hosted in the EU

• Reads your chats: No
• Verify yourself: Yes
• Trains by default: No
• Models: 3 chat models¹⁵
• Price: 5M free tokens, then per token
• Platforms: Web, API

Privatemode, from Edgeless Systems, runs models in confidential VMs and GPUs. Its client proxy or SDK verifies the deployment before sending, and builds are reproducible. Prompts "aren't stored after a request completes, and they aren't used for training," and are processed exclusively within the EU, according to its trust and compliance page.

Edgeless holds ISO 27001, and Privatemode meets the BSI C5 attestation criteria. The free tier gives 5M tokens at sign-up plus 1M a month.

Where it falls short: there are no mobile apps, the chat app is basic, and metadata such as IP address and token usage is kept for up to 90 days. The code is source-available, and the hosted web app adds Edgeless and a network provider to what you trust.¹⁶

Best suited for: EU companies that want attested inference billed per token, processed only in the EU.

Checked 15 Oct 2026.

Apps often called private, and what they protect

These apps appear on most private AI lists. Each protects something real, but in each one the company or a model provider can read your prompts, and there's nothing for you to check. We don't rank them.

Proton Lumo

Lumo stores your chat history with zero-access encryption, so Proton can't read your saved chats. That encryption covers history, not processing. Proton's security model post says "the LLM gets to see the decrypted user message and process it directly," and that the server forgets the request once the response is generated.

Proton runs the models on servers it controls in Europe and says on its Lumo privacy page that it keeps no logs of prompts or replies and doesn't train on your chats. Lumo Plus costs $12.99 a month.

Checked 15 Oct 2026.

Venice

Venice labels each model with one of four modes on its privacy page. With "Anonymized" models, the provider "sees and likely saves your prompt." The default "Private" mode runs on Venice or partner GPUs, which Venice calls "Trust-based."

The "TEE" and "E2EE" modes on Pro use enclaves verified by remote attestation, which Venice's TEE docs expose through the API; 12 text models run in E2EE mode²⁷. Its privacy policy doesn't address training on chat content.

Checked 15 Oct 2026.

Duck.ai

DuckDuckGo removes metadata such as your IP address and forwards prompts to OpenAI, Anthropic and together.ai, per its page on what Duck.ai shares. So the provider sees your text, but not who sent it. Two open models are processed in a trusted execution environment, which DuckDuckGo doesn't let you verify.

Its privacy terms say providers don't train on prompts and delete them within 30 days, with limited exceptions; its help page says right after replying, with exceptions. Paid plans start at $9.99 a month.

Checked 15 Oct 2026.

Brave Leo

Leo proxies requests "through an anonymized server so the request and user-address cannot be linked," per Brave's launch post. Brave's browser privacy policy says "We don't store or retain prompts, responses, context, or personal data on our servers."

For most models, whoever runs inference still sees your prompts, and the policy doesn't name those hosts. Leo Premium costs $14.99 a month.

In October 2026, Brave added end-to-end encrypted chat with one open-weight model, GLM 5.3 Flash, in beta. The browser encrypts each prompt to an enclave run by NEAR AI on Intel TDX and NVIDIA confidential-computing hardware, and Brave's relay can't read it. Today Brave's gateway checks the enclave's attestation, not your browser; Brave says an in-browser check comes next, and its verifier and relay code is open source.

Checked 20 Oct 2026.

ChatGPT and Claude with privacy settings

OpenAI's consumer data page says authorized OpenAI staff and trusted service providers can access content to investigate abuse or security incidents, provide support, handle legal matters, and improve its models unless you opt out. OpenAI may use chats on its personal plans (Free, Go, Plus and Pro) to train its models unless you turn that off in data controls.⁴⁰

Even after you opt out, giving a response a thumbs up or down lets OpenAI use that whole conversation for training. OpenAI may keep a temporary chat for up to 30 days.

Please do not enter sensitive information that you would not want reviewed or used.

Anthropic uses your chats for training if you choose to allow it, and keeps incognito chats for 30 days by default. These settings change what happens to your chats, not who can read them; more in Is ChatGPT private?

Checked 15 Oct 2026.

Also checked

• Meta AI Incognito Chat. It runs on Meta's Private Processing system, an attested enclave reached through a third-party relay,⁴⁴ and Meta says "no one — not even Meta — can read your conversations."⁴⁵ It works only inside WhatsApp and the Meta AI app, and those apps are closed source.
• Kagi Assistant. It sends prompts to third-party model providers,⁴⁶ under a policy that bars training and sharing your account details, and deletes threads after 24 hours of inactivity by default.⁴⁷ It is mainly part of a search subscription.
• Mistral Vibe, formerly Le Chat. Mistral uses input and output to train its models by default unless you opt out.⁴⁸

If you'd rather run it yourself

• Reads your chats: No one; it runs on your device
• Verify yourself: Ollama and Jan are open source
• Trains by default: No
• Models: Any open-weight model your hardware can run
• Price: Free
• Platforms: macOS, Windows, Linux

A local app runs the model on your own computer, so there's no company to trust. LM Studio says "none of your messages, chat histories, and documents are ever transmitted from your system." Ollama and Jan say the same of local use, and both are open source.

The trade-offs are hardware and model size. LM Studio recommends 16GB or more of RAM, models that fit on a laptop are smaller than large hosted ones, and your history is only as safe as your device. If you need everything fully offline, it's the right tool.

Checked 15 Oct 2026.

Side-by-side comparison

Only the first three rows pass both tests. Platforms for those three apps are in each entry above.

Each row's sources are linked in that app's section above.

Which one fits you

To go further, our guide to using AI privately covers your connection, email and payment too. For clinical work, see using AI safely in a therapy practice.

Check it yourself

A no-logs promise can't be checked from outside. An attestation can. When you open Maple, the app asks the enclave for a signed attestation document and checks four things: a fresh random challenge, a certificate chain back to the enclave platform's published root, a valid signature, and an enclave image fingerprint that matches a release Maple approved. If any check fails, nothing is sent.

You can run the same check in your browser, and the repository documents how to rebuild the enclave image and compare its fingerprint with the public source. The proof page walks through each step.

Try Maple free

The quickest way to judge a private AI app is to use one and check it. Maple has a free plan, and no credit card is required to sign up. Your chats are encrypted on your device and decrypted only inside a hardware-isolated enclave, and the app checks that enclave before it sends anything.

The same account works on web, iPhone, Android, Mac, Windows and Linux. You can also sign up without an email and pay with Bitcoin. Pro, at $20 a month, adds every model, image and document upload in the apps, and API access. Start a free Maple chat, then run the check yourself on the proof page.

Sources

1. Maple, How Maple works: private AI you can check. Enclave design, attestation checks, training and audit status.
2. Maple, Pricing. Plans, Free plan features and models.
3. Maple, Download Maple. Platforms.
4. Maple, Maple source code. MIT license, enclave service and approved image fingerprints.
5. Confer, Confer: Private AI. In-app description: models, platforms, Guest plan and Membership.
6. Confer, Private inference, January 2026. Attestation, reproducible builds and transparency log.
7. Confer, Privacy policy, October 2026. Encryption, stored history and training.
8. Apple App Store, Confer Private AI. Membership price, iOS release and platforms.
9. Confer, ConferLabs on GitHub. Published server repositories: one Apache-2.0, one with an AGPL-3.0 component, one without a license.
10. Confer, Confer is bringing foundational AI privacy to Meta, March 17, 2026.
11. Confer, Confessions to a data lake, December 23, 2025. First public release on the web.
12. Privatemode, Attestation overview. Client-side verification and reproducible builds.
13. Privatemode, Trust and compliance. Retention, training, EU processing, metadata and certifications.
14. Privatemode, Pricing. Free tokens and per-token pricing.
15. Privatemode, Models.
16. Privatemode, Web app. What the hosted web app adds to the trusted computing base.
17. LM Studio, Privacy policy. Local use and cloud services.
18. LM Studio, System requirements. RAM recommendation.
19. Ollama, Privacy policy. Local use.
20. Jan, Privacy. Local storage and remote APIs.
21. Proton, Lumo security model: how Proton makes AI private. Encryption in transit and decryption on Proton's servers.
22. Proton, Lumo privacy. No-logs policy, training, models and hosting.
23. Proton, Lumo pricing and plans.
24. Venice, Privacy in Venice. The four privacy modes and local history.
25. Venice, TEE and E2EE models. Attestation through the API.
26. Venice, Privacy policy, effective June 2, 2026.
27. Venice, All models. Model counts by mode.
28. Venice, Pricing. Free plan and Pro price.
29. DuckDuckGo, What information does Duck.ai share with model providers?. Providers and enclave-processed models.
30. DuckDuckGo, Duck.ai privacy policy and terms of use. Retention, training and local history.
31. DuckDuckGo, How does Duck.ai protect my privacy?. Provider deletion and its exceptions.
32. DuckDuckGo, DuckDuckGo's Pro plan, March 18, 2026. Plus and Pro prices.
33. Brave, Leo, Brave's AI browser assistant. The anonymizing proxy.
34. Brave, Browser privacy policy, updated September 23, 2026. Retention and training.
35. Brave, Brave Premium. Leo Premium price.
36. Brave, Leo adds end-to-end encryption, Oblivious HTTP, and verifiable confidential computing, October 15, 2026. The encrypted GLM 5.3 Flash mode, its hardware and who checks the attestation.
37. OpenAI, How OpenAI handles data in consumer services. Who can access chats and why.
38. OpenAI, How your data is used to improve model performance. Training, opt-out and feedback.
39. OpenAI, Temporary chat in ChatGPT. Training and 30-day retention.
40. OpenAI, ChatGPT pricing. Plan prices and the training opt-out on personal plans.
41. Anthropic, Is my data used for model training?.
42. Anthropic, Use incognito chats.
43. Anthropic, Claude plans and pricing.
44. Meta, Building Private Processing for AI tools on WhatsApp, April 29, 2025. Attested enclave and third-party relay.
45. Meta, Introducing a completely private way to chat with AI, May 13, 2026.
46. Kagi, LLMs and privacy. Model providers.
47. Kagi, Kagi Assistant. Training, account sharing and thread deletion.
48. Mistral AI, Do you use my user data to train your artificial intelligence models?.

Frequently Asked Questions

Is there an AI that is truly private?

Yes. Three hosted apps are built so the company can't read what you type, and let you check that yourself: Maple, Confer and Privatemode. A model running on your own computer also keeps your prompts private. Most other apps called private rely on a policy you have to trust.

What is the most private AI chat app?

Maple, Confer and Privatemode are the hosted apps the company can't read and you can verify, and Maple is the only one of them with apps on every major platform, an API and a free plan. If nothing may leave your device, a local app such as LM Studio, Ollama or Jan keeps prompts on your own computer.

How private is Claude compared with ChatGPT?

Neither one is built so the company cannot read your chats; both rely on policy. OpenAI may use chats on its personal plans (Free, Go, Plus and Pro) for training unless you opt out, and Anthropic trains on chats if you choose to allow it in your settings. ChatGPT's temporary chats and Claude's incognito chats are not used for training but can be kept for up to 30 days.

How can I make ChatGPT more private?

You can make it more private: turn off Improve the model for everyone and use Temporary Chat, which OpenAI may still keep for up to 30 days. OpenAI still receives your chats in a form it can read. For chats the company can't read, use an attested enclave app or a local model.

Can the company or my employer see what I type into an AI chat?

Whoever runs the model can read your prompt unless it runs in an attested enclave or on your own device. Anonymizing apps such as Duck.ai hide who you are from the model provider, but the provider still receives the text. On a work account, your employer's plan and settings decide what administrators can see, so ask your IT team.

Is Proton Lumo really confidential?

Lumo stores your chat history with zero-access encryption, so Proton cannot read your saved chats. Each new prompt is decrypted on Proton's servers so the model can answer it, and Proton says it keeps no logs of what you ask. That is a policy promise, because there is no attestation for you to check.

What does a private AI chat app cost?

Most have a free tier. Paid plans run from $8 a month for ChatGPT Go to $34.99 for Confer; Maple Pro is $20, ChatGPT Plus and Claude Pro are $20, and Privatemode bills per token. Local apps are free, but you pay in hardware.

How can I run my own private AI?

Yes. LM Studio, Ollama and Jan run open-weight models on your own computer for free, and your prompts stay on the device while you use a local model. LM Studio recommends 16GB or more of RAM, and models that fit on a laptop are smaller than the large hosted ones.

It keeps them until you delete them, can train its models on them unless you turn that off, and lets authorized staff and contractors read some of them. Settings change how OpenAI uses your chats. None of them changes who can read them.

Every answer below comes from OpenAI's own pages and from court records. After that, we compare ChatGPT with six other AI chat services on the same questions, including one most of them can't answer yes to: can you check the claim yourself?

Can anybody see my ChatGPT chats?

Yes, in specific cases. OpenAI's automated systems scan chats for abuse, and a limited number of OpenAI staff and contractors can read them for stated reasons. Anyone you send a share link to can read that chat, and on a work account your employer's admins can access it. Courts and police can reach chats through legal process.

OpenAI lists four reasons its personnel and service providers may access your content: investigating abuse or a security incident, answering a support request, handling legal matters, and improving model performance unless you've opted out.¹ It logs that access and requires training first. Its own advice is blunt:

Please do not enter sensitive information that you would not want reviewed or used.

Flagged conversations are assessed by trained personnel. When a chat indicates "an imminent and credible risk of harm to others," OpenAI says it notifies law enforcement.² Our guide to whether AI companies can report you to the police covers how that review and referral works across providers.

Work accounts add another reader. OpenAI's privacy policy says the administrators of a ChatGPT Enterprise or business account "may access and control your OpenAI account, including being able to access your Content."³

For US law enforcement, OpenAI's policy since January 1, 2026 says it discloses chat content only in response to a valid warrant or equivalent, and that it tells the user first unless the law prohibits it.⁴ Civil courts are a separate route, covered below.

Does ChatGPT use your chats for training?

Yes, unless you turn it off. OpenAI says it may use your content from ChatGPT to train its models. To stop that for new conversations, turn off Improve the model for everyone under Settings > Data controls. A thumbs up or thumbs down still sends that whole conversation to training.

OpenAI's help page states the feedback exception plainly: "If you choose to provide feedback, the entire conversation associated with that feedback may be used to train our models."⁵ It also says it takes steps to reduce the personal information in training data first. Opting out applies to new conversations only.

Memory follows the same switch. For personal accounts, OpenAI may use chats and remembered information to improve models when the setting is on.⁶ ChatGPT Business, Enterprise, Edu and the API are not used for training by default.

How long does ChatGPT keep your chats?

Until you delete them. Regular and archived chats stay in your account with no automatic expiry. When you delete one, OpenAI schedules it for permanent deletion within 30 days, unless it was already de-identified or OpenAI must keep it longer for security or legal reasons, such as a subpoena.⁷

Two details catch people out. Deleting a chat doesn't necessarily delete a memory ChatGPT saved from it, and deleting a chat doesn't delete files you saved to your Library.

The legal exception is not hypothetical. In May 2025, in The New York Times' copyright case, a federal court ordered OpenAI to preserve output logs that would otherwise be deleted, including at a user's request.⁸ That obligation ended on September 26, 2025, but logs preserved before then are still held, apart from requests from the EEA, Switzerland and the UK.⁹

In January 2026, Judge Sidney Stein affirmed orders requiring OpenAI to produce 20 million de-identified ChatGPT logs to the plaintiffs.¹⁰ He compared them with wiretapped phone calls and found the privacy interest weaker:

…the privacy interests in users' conversations with ChatGPT which users voluntarily disclosed to OpenAI and which OpenAI retains in the normal course of its business.

The timeline below shows how much changed in nine months.

Is ChatGPT's Temporary Chat private?

It's more private than a regular chat, but OpenAI can still read it. A temporary chat stays out of your history, isn't used for training while it stays temporary, and doesn't create new memories. OpenAI may keep a copy for up to 30 days for safety, and longer if it must for safety or legal reasons.

OpenAI's help page now describes two newer behaviors.¹¹ You can save a temporary chat, which turns it into a regular chat under your normal training setting. You can also let it use your existing memories. Temporary chats don't show ads.

Is ChatGPT encrypted?

Yes, in transit and at rest, but not end to end. OpenAI says your content "is encrypted at rest and in transit between you and OpenAI, and between OpenAI and its service providers."¹² That protects chats from someone intercepting the connection or stealing a disk. It doesn't keep them from OpenAI.

With end-to-end encryption, only the people in a conversation can read it. No OpenAI page claims that for ChatGPT, and its staff-access policy rules it out. OpenAI applies the encryption itself, so its systems can decrypt your chats to answer them and to let authorized staff review them.

Does ChatGPT share your data with advertisers?

OpenAI says it doesn't share your conversations with advertisers or sell your data. Since February 9, 2026, though, it has been testing ads on the Free and Go plans, starting in the US, and it picks them using your current chat. With ad personalization on, past chats and memory count too.

Plus, Pro, Business, Enterprise and Edu have no ads.¹³ OpenAI says ads don't appear near sensitive topics, including personal health, mental health and politics. Turning off ad personalization stops the use of other chats, but ads still follow the context of the chat you're in.

How do ChatGPT, Claude, Gemini and other AI chats compare?

They differ on training, review and retention, but every major chat app receives your regular chats in a form the company can read. The table compares consumer plans as each company describes them on its own pages. "Not stated" means we found no statement, not that the control is missing.

Some of these protections deserve credit. Anthropic says its employees can't access your conversations by default, except feedback you send and chats its Trust & Safety team reviews to enforce its usage policy.¹⁹ Claude's incognito chats and Grok's Private Chat stay out of training, like ChatGPT's Temporary Chat, and both are still kept for up to 30 days.

Gemini keeps more, for longer. Even with Keep Activity off, chats are saved for 72 hours, and chats a human reviewer has read are kept for up to three years, disconnected from your account, even after you delete your activity.

Google and Meta both say a few features run where even they can't read your data: Google's Private AI Compute for Pixel features such as Magic Cue and Recorder summaries, and Meta's Incognito Chat with Meta AI.²⁰²¹ You can't verify either claim. Neither gives you a check you can run, neither publishes all of the code it runs there, and neither covers regular chats in the Gemini app or Meta AI.

Their own auditor says as much. NCC Group, hired by both companies, wrote that Meta's users "must trust Meta" not to slip malicious behavior into code that isn't open source.²² Of Google's system, it wrote that "it is logically unavoidable that Google ultimately has that power."²³

Can you check any of this yourself?

Not for ChatGPT, or for any of the other major apps. Each publishes a privacy policy, and you trust that its staff and systems follow it and that it won't change in ways you'd object to. Policies can be good and sincerely meant. You still can't inspect the servers that carry them out.

The alternative is to make the claim checkable. A secure enclave is a hardware-isolated part of a server that the people running it can't look inside. An attestation is a signed document stating exactly which software is running there, so your app can confirm the code before it sends anything.

That's how Maple works. Before Maple's app sends anything, it asks the enclave for a signed fingerprint of its software and compares it with values we've approved. If any check fails, nothing is sent. You can run the same check in your browser and watch each step.

The model runs in its own attested enclave too. Maple's enclave verifies it before forwarding a prompt, and the code that does so is part of the software you check: you check Maple's enclave, and Maple's enclave checks the model's. You still rely on the hardware behind each enclave, the published code and Maple's signing key. Our proof page names all of them.

How do you make ChatGPT more private?

Change five settings and habits. They limit how OpenAI uses your chats. They don't change the fact that OpenAI receives them in a form it can read, so the most effective step is still the last one: leave out what you wouldn't want a reviewer to see.

1. Turn off Improve the model for everyone under Settings > Data controls.
2. Use Temporary Chat for anything you don't want in your history, and skip the thumbs up and down buttons on chats you want kept out of training.
3. Review saved memories and delete the ones you don't want, since deleting a chat may not delete them.
4. Delete old chats and shared links you no longer need. On Free and Go, turn off ad personalization if you don't want past chats used for ads.
5. Keep personal topics off a work account, where your employer's admins can access your chats.

If you also want to hide your IP address or your real email from an AI service, our guide on how to use AI privately covers those layers. A VPN, for example, hides where you connect from but not what you say.

Is there a version of ChatGPT that's private?

Not from OpenAI. ChatGPT Business and Enterprise aren't used for training by default, and OpenAI narrows when its staff can access Enterprise chats. But your employer's admins can access them, and OpenAI still receives them in readable form. Privacy from the company itself takes a service built so the operator can't read your chats.

That's the version we built. Maple gives you a ChatGPT-style assistant on open models, without running your own hardware, and no one at Maple can read what you say. Our model library lists the models you can use.

If your work involves client or patient information, privacy tools don't replace your professional duties. Our guide to using AI safely in a therapy practice covers how those rules apply.

Sources

1. OpenAI, How OpenAI handles data in consumer services, updated September 2026; checked October 15, 2026. Who can access chats and why, and access logging.
2. OpenAI, Our commitment to community safety, April 28, 2026. Review of flagged conversations and notifying law enforcement.
3. OpenAI, Privacy Policy, updated September 10, 2026; checked October 15, 2026. Workspace administrators, retention exceptions and ads.
4. OpenAI, Government User Data Request Policy, effective January 1, 2026. Warrant requirement for content, user notice and emergency disclosure.
5. OpenAI, How your data is used to improve model performance, updated September 2026; checked October 15, 2026. The training setting, the feedback exception and business defaults.
6. OpenAI, Memory FAQ, updated September 2026; checked October 15, 2026. Memories, deletion and training.
7. OpenAI, Chat and file retention in ChatGPT, updated July 2026; checked October 15, 2026. How long chats are kept and the 30-day deletion schedule.
8. U.S. District Court for the Southern District of New York, Magistrate Judge Ona T. Wang, Order on preservation of output log data, The New York Times Co. v. Microsoft Corp. et al., No. 23-cv-11195, May 13, 2025.
9. U.S. District Court for the Southern District of New York, Stipulation and Order to Terminate OpenAI's Ongoing Obligations Under the Preservation Order, No. 23-cv-11195, October 9, 2025.
10. U.S. District Court for the Southern District of New York, Judge Sidney H. Stein, Order affirming discovery orders on ChatGPT output logs, In re OpenAI, Inc. Copyright Infringement Litigation, No. 25-md-3143, January 5, 2026.
11. OpenAI, Temporary chat in ChatGPT, updated September 2026; checked October 15, 2026.
12. OpenAI, Security & privacy, checked October 15, 2026. Encryption in transit and at rest.
13. OpenAI, Ads in ChatGPT, updated September 2026; checked October 15, 2026. Plans with ads, how ads are chosen and what is shared with advertisers.
14. Anthropic, How long do you store my data?, July 1, 2026; checked October 15, 2026. Deletion, training retention and incognito chats.
15. Google, Gemini Apps Privacy Hub, updated October 13, 2026; checked October 15, 2026. Keep Activity, human review, 72-hour and three-year retention.
16. Microsoft, Microsoft Copilot for individuals: your activity history, August 18, 2026, with the Microsoft Privacy Statement, September 2026; checked October 15, 2026.
17. Meta, AI Terms of Service, effective May 13, 2026, with the Meta Privacy Policy, effective July 23, 2026; checked October 15, 2026. Review, retention and use of AI chats.
18. xAI, Privacy Policy, updated October 11, 2026, with the consumer FAQ; checked October 15, 2026. Training setting, Private Chat, human review and deletion.
19. Anthropic, How does Anthropic protect the personal data of Claude users?, March 16, 2026; checked October 15, 2026. Employee access and encryption.
20. Google, Private AI Compute, November 11, 2025. Remote attestation for certain Pixel features.
21. Meta, Introducing Incognito Chat with Meta AI, May 13, 2026. What Incognito Chat promises and that it runs on Private Processing.
22. NCC Group, Public report: Meta WhatsApp Message Summarization Service review, August 27, 2025. Commissioned by Meta; covers the Private Processing summarization service.
23. NCC Group, Public report: Google Private AI Compute review, version 1.4, November 8, 2025. Commissioned by Google.

Frequently Asked Questions

Can people see what I type into ChatGPT?

OpenAI can. Your chats reach its servers in a form it can read, its automated systems scan them for abuse, and a limited number of authorized staff and contractors can read them for stated reasons, including model improvement unless you opt out. Other people see a chat only if you share a link to it, use a work account your employer administers, or a court or police reach it through legal process.

Is ChatGPT safe for confidential information?

Not on consumer plans, by OpenAI's own advice: it asks you not to enter sensitive information you would not want reviewed or used. Business and Enterprise plans skip training by default, but OpenAI still receives the chats in readable form and your employer's admins can access them. For client or patient information, check your professional duties first; a privacy setting does not replace them.

Can my employer see my ChatGPT history?

On a work account, yes. OpenAI's privacy policy says the administrators of a ChatGPT Enterprise or business account may access and control your account, including your content. On a personal account your employer has no access through OpenAI, though a work device or network can be monitored separately.

Does ChatGPT share your data with the government or police?

Only in specific cases. For US law enforcement, OpenAI's policy requires a valid warrant for chat content and a subpoena or court order for other account data, and it notifies the user first unless the law prohibits it. It can also disclose information in an emergency involving danger of death or serious injury, and it says it notifies law enforcement when a chat shows an imminent and credible risk of harm to others.

Which AI chat is the most private?

It depends on what you mean by private. Among the major apps, Claude restricts employee access by default and Gemini keeps reviewed chats the longest, but all of them can read your regular chats. If you want a service whose operator cannot read your chats, look for one built on attested hardware that lets you check the claim, or run a model on your own computer.

There are three broad ways to use it. You can run the weights yourself, which takes a multi-GPU server. You can call DeepSeek's API or another host. Or you can use it in Maple, where we run the 1017 build on Pro, Max and Team plans with the full 1,048,576-token window, through our protected inference path.

This guide separates what belongs to the model from what belongs to each route, so you can pick one.

What is DeepSeek V4.1 Flash?

It is a sparse model: of its 384 routed experts per layer, six run for each token, plus one shared expert. DeepSeek splits the network into a causal encoder that reads your prompt and a decoder that writes the answer, which is why it quotes two active counts: 8B in prefill and 16B in decode. That design keeps the compute per token low, especially for long inputs.

It does not shrink the model you have to store. Every expert has to be loaded, because different tokens route to different experts. The card also describes 196B parameters of "Engram" conditional memory, a lookup table the model reads as it works, which sits beside the backbone.

The card's headline change is the cache that long contexts need. DeepSeek reports a global KV cache of 890 bytes per token, about a quarter of V4 Flash's. By our arithmetic from that figure, a full million-token window comes to under a gigabyte of global cache per conversation, before the other buffers a serving engine adds. For this model the weights dominate the memory, not the context.

Is DeepSeek V4.1 Flash open source?

The weights are. DeepSeek publishes them under the MIT License, together with reference inference code and a technical report. MIT is a standard permissive license, unlike the custom licenses some open-weight models use, so building a product on the weights does not bring revenue or service conditions.

What DeepSeek does not publish is its training data. If your definition of open source requires the data, V4.1 Flash is open-weight rather than fully open.

How is V4.1 Flash different from V4 Flash?

They are separate releases, and V4.1 Flash replaced V4 Flash on DeepSeek's API. DeepSeek's release note retired V4 Flash and its vision preview on launch day. For compatibility, requests that still name deepseek-v4-flash now get V4.1 Flash. So a test that compares the two by API name may be comparing V4.1 Flash with itself.

The figure below puts the API changes in order. DeepSeek also routes deepseek-v4-pro requests to V4.1 Flash until a V4.1 Pro arrives, which matters if you built on V4 Pro.

What changed in the 1017 update?

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17 and calls it a minor update to V4.1 Flash, published as new weights in their own repository. The model you store doesn't change. It is the same 48 weight files and 510.3 GB, under the same MIT License, with the same million-token window.

What changes is the training. DeepSeek reports stronger reasoning, coding, agent and visual results, plus shorter answers that finish tasks faster. On DeepSeek's API, deepseek-flash now serves 1017, and DeepSeek says calls and pricing are unchanged.⁸

The largest gains are in agentic coding, the long, multi-step work in a terminal or across a repository. Read the Terminal-Bench row with care. The September card ran it in DeepSeek's own harness, the 1017 card names a different one, and both list the same September score.

What do DeepSeek's evaluations show?

They show strong vendor-reported results. DeepSeek's technical report explains the architecture behind them,⁹ and the model card compares V4.1 Flash with V4 Flash, V4 Pro and several frontier models, and DeepSeek's release note says it scores ahead of V4 Pro. These are DeepSeek's results: it chose the benchmarks, the settings and the baselines, and they are not Maple measurements.

The settings matter more than usual here. DeepSeek ran its instruct results at the maximum of a reasoning-effort scale that goes from 1 to 100, and its coding-agent results with its own harness and a one-million-token window. Another host or app may run the model at a lower effort or a shorter window, and get different results.

A few habits help when reading the tables:

• Match the settings. A score earned at maximum effort with a 1M window says little about a quick answer at default settings.
• Check the harness. The card itself shows the same model scoring differently across agent scaffolds on the same benchmark.
• Treat benchmarks as a shortlist. A strong coding result is a reason to try V4.1 Flash on coding work, not a promise about your codebase.

The most reliable evidence is still a test on your own representative tasks.

What hardware does DeepSeek V4.1 Flash need to self-host?

A multi-GPU server, in every published recipe we found. Running the weights yourself gives you the most control over the model, the serving stack and where your data goes. The cost is hardware and engineering time at a scale few teams have spare.

We checked the download size ourselves. Hugging Face's file listings for the September and 1017 repositories each show 48 weight files totaling 510.3 GB.⁵ The listing's own parameter total, 763B, counts stored values in mixed formats, so it does not match DeepSeek's 552B backbone and 196B Engram figures; read DeepSeek's numbers for the model and the file size for your disk.

The vLLM recipe is the clearest public serving guide, updated October 17, 2026, when it switched its default checkpoint to 1017. It separates three budgets that are easy to confuse:

That third budget is what makes the smaller setups work. The recipe calls a node with eight 80 GB H100 GPUs "the smallest NVIDIA node that serves this checkpoint", and only with the Engram tables moved to host memory and other memory limits applied. It also documents a single DGX Station GB300 that moves the Engram tables and 60 GiB of experts to host memory and still serves the full 1M window.

Those results belong to the recipe's configuration. On NVIDIA it installs from the vLLM 0.31.0 release; some AMD setups still use a nightly build. Neither the recipe nor DeepSeek's card describes a consumer GPU setup. Your own needs depend on the checkpoint format, context length, concurrent users and serving engine.

For an open model that runs on a single GPU, see our Gemma 4 31B guide.

Self-hosting is the right call when:

• your information must never leave hardware you control
• you need to modify, fine-tune or deeply instrument the model
• you already run multi-GPU inference and have capacity to spare

If none of those apply, ask whether the operational work buys you anything a hosted route can't. Our Kimi K3 guide works through the same question for a larger open-weight model.

How do you use DeepSeek V4.1 Flash in Maple?

Choose DeepSeek V4.1 Flash from the model picker in Maple Research on a Pro, Max or Team plan. It runs the 1017 build, under the same name. It is not available on Free, where GPT-OSS 120B is the model to start with. We configure it with the full 1,048,576-token window, so a long contract set, a large codebase or months of notes can go into one conversation.

Plan details are on our pricing page. DeepSeek V4.1 Flash also appears in our model library with the other models we offer; our GLM-5.3 Flash guide covers the other model there with a million-token window.

You can upload an image while using DeepSeek V4.1 Flash in Research. The app first has a vision model describe the image, then gives that description to the model you selected.¹⁰ DeepSeek V4.1 Flash answers from the description, not a direct view of the image, which matters if your task depends on a small visual detail.

For developer tools and existing applications, Maple Proxy is our API route, using the model ID deepseek-v4-1-flash. Check the Proxy model list for its current context limit and input types before you build on it.

This route gives you two things. You get a model with a 510 GB checkpoint without provisioning the GPU memory above or tracking serving builds. And your requests run through our protected inference path rather than a conventional hosted API.

The route has limits to plan around:

• Images are described, not seen. In Research, the model works from a vision model's description of each image.
• A paid plan is required. DeepSeek V4.1 Flash is on Pro, Max and Team.
• Long prompts take time. A window near a million tokens is slow to process on any route, ours included.

What information reaches each route?

That depends on who operates the model: you, a conventional host, or Maple. Three questions get blurred together here. Who can read your prompts while they are processed, what happens to them afterward, and is any of it used for training?

Self-hosted weights. Your prompts and outputs stay on infrastructure you operate. The trust boundary is your own: your access controls, logging and retention. That makes this the strongest option when information must stay on your hardware, and it makes you responsible for securing it.

DeepSeek's API or another host. DeepSeek's API terms say it handles a developer's personal information under its privacy policy.¹¹ That policy lists prompts and uploaded files among the data it collects, and says DeepSeek stores personal data in the People's Republic of China.¹²

It is a clear disclosure. Other hosts that serve the open weights set their own terms, so read them for the service and plan you would use.

Maple. We run DeepSeek's open weights inside an attested enclave rather than calling DeepSeek's API. Your requests are encrypted before they leave your device, and our proof page explains the attestation you can check yourself.

The model runs in its own attested enclave, and Maple's enclave verifies it before forwarding a prompt. It works like certificate chains on the web: you check Maple's enclave, and Maple's enclave checks the model's. Our privacy notice covers everything outside your chats.

That design has a scope. It protects the contents of your chats: your prompts and the responses. It doesn't make information private if you've already shared it elsewhere.

It is also separate from the safety of what the model's output leads you to do. If you paste its code into production or let an agent act on its suggestions, confidential processing doesn't vouch for that code or those actions. Review outputs as you would any untrusted input.

Which route fits your task?

It depends on where your information may go, how much context you need and how much infrastructure you can run. The table compares the three routes on the same terms.

A rough decision rule:

• Choose self-hosting if information must stay on your hardware, or you need control over the weights themselves.
• Choose DeepSeek's API or another host if your prompts aren't sensitive and the provider's terms fit your data.
• Choose Maple if you want the full million-token window on sensitive material without running a multi-GPU server.

Sources

1. DeepSeek-AI, DeepSeek-V4.1-Flash model card, last updated October 16, 2026, accessed October 18, 2026. Source for the architecture, parameter and expert counts, Engram memory, context length, input types, KV cache figures and evaluation settings.
2. DeepSeek-AI, MIT License for DeepSeek-V4.1-Flash, accessed October 15, 2026.
3. DeepSeek-AI, DeepSeek-V4.1-Flash-1017 model card, October 17, 2026, accessed October 18, 2026. Source for the "minor update" description, the benchmark comparison with the September build and the shorter outputs.
4. DeepSeek, “DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient”, September 10, 2026. Source for the release, the deepseek-flash API name, the retirement of V4 Flash, the routing of deepseek-v4-flash and deepseek-v4-pro, and DeepSeek's comparison with V4 Pro.
5. Hugging Face, repository file listings for DeepSeek-V4.1-Flash and DeepSeek-V4.1-Flash-1017, checked October 18, 2026. Our sums of the 48 safetensors files in each (510.3 GB) and the listing's stored-value total.
6. vLLM project, DeepSeek-V4.1-Flash serving recipe (source file), updated October 17, 2026, accessed October 18, 2026. Source for the 614 GB default minimum, the verified hardware, the H100 and DGX Station GB300 configurations, the 183 GiB Engram tables, the 1017 default and the vLLM 0.31.0 requirement.
7. Maple, public model configuration and Research model picker, checked October 19, 2026, and the routing change that serves the 1017 checkpoint on every DeepSeek route. Source for the plan tier, configured context window, listed input types and the 1017 build in Maple.
8. DeepSeek, API change log, entry of October 17, 2026, accessed October 18, 2026. Source for deepseek-flash serving 1017 with unchanged calls and pricing.
9. DeepSeek-AI, “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression”, arXiv, submitted September 17, 2026. DeepSeek's technical report.
10. Maple, Research image handling and vision helper, checked October 19, 2026. Source for image description before the selected model responds.
11. DeepSeek, Open Platform Terms of Service, effective April 29, 2026, accessed October 15, 2026. Section 5.5 applies the privacy policy to a developer's personal information.
12. DeepSeek, Privacy Policy, last updated February 10, 2026, accessed October 15, 2026. Source for the prompts and files DeepSeek collects and where it stores personal data.

Frequently Asked Questions

Is DeepSeek V4.1 Flash open source?

The weights are published under the MIT License, a standard permissive license, so you can download, modify and build on them. DeepSeek also publishes inference code and a technical report. It does not publish its training data.

Does Maple support DeepSeek V4.1 Flash's full 1M-token context?

Yes. We configure DeepSeek V4.1 Flash with a 1,048,576-token window, the same as the upstream maximum. Very long prompts take longer to process on any route.

Does Maple run the DeepSeek V4.1 Flash-1017 update?

Yes. Maple serves DeepSeek's October 17 build, DeepSeek-V4.1-Flash-1017, under the same name in the model picker and the same deepseek-v4-1-flash ID on Maple Proxy, so there is nothing to change on your side.

Can I upload an image with DeepSeek V4.1 Flash in Maple?

Yes. In Maple Research, a vision model first describes an uploaded image, and DeepSeek V4.1 Flash works from that description, not the image itself. Plan for that if your task depends on a small visual detail.

What hardware do I need to run DeepSeek V4.1 Flash locally?

The download is about 510 GB. The vLLM recipe's default setup needs at least 614 GB of GPU memory, and its smallest NVIDIA option is a node with eight 80 GB GPUs that keeps part of the model in host memory. No recipe we found targets a consumer GPU.

Is deepseek-v4-flash on DeepSeek's API still the older model?

No. DeepSeek retired V4 Flash on September 10, 2026. Requests that name deepseek-v4-flash now route to V4.1 Flash for compatibility, and deepseek-v4-pro requests have also routed to V4.1 Flash since September 14.

Can Maple read my DeepSeek V4.1 Flash chats?

No. No one at Maple can read your chats. They are encrypted on your device and decrypted only inside a hardware-isolated enclave running open-source code you can verify, and your chat history is stored encrypted. Our privacy notice covers everything outside your chats.

There are three broad routes. You can run the weights yourself, which requires very large serving infrastructure. You can use Moonshot's API or another host. Or you can use K3 in Maple, where we offer it on our Pro, Max, and Team plans with a 262,144-token context window, through our protected inference path.

Each route differs in context length, input types, setup burden, and where your information goes. This guide separates what belongs to the model from what belongs to each route so you can choose among them.

Sources were checked on 11 October 2026. Model support, plans, and deployment recipes can change.

What is Kimi K3?

Kimi K3 is a mixture-of-experts model from Moonshot AI with 2.8T total parameters, of which 104B are activated for each token. The gap between those two numbers comes from the design: the model contains many specialized sub-networks, and only some of them run for each token. That reduces the computation per token compared with a dense model of the same total size.

It does not reduce the amount of model you have to store. All 2.8T parameters must be available to the serving system, because different tokens route to different experts. A simple calculation shows the scale: at one byte per parameter, the weights alone come to roughly 2.8 terabytes; at half a byte, roughly 1.4 terabytes.

Those figures cover only the weights. They leave out the memory a long context consumes during inference and any serving overhead, and they are not a hardware recommendation for any particular quantization.

The card's specification table also lists a 1,048,576-token context window, roughly a million tokens, with text and image input. The card's overview says K3 understands video as well, but its specification table lists only text and image. Confirm video support on the route you plan to use.

Is Kimi K3 open source?

Not in the standard sense. K3 is open-weight: Moonshot publishes the weights for download under the Kimi K3 License, a custom license. It is not MIT or another standard open-source license, and it includes conditions related to offering services and to revenue.

If you plan to build a product or service on the weights, read the license yourself and get legal advice where it matters. This guide describes the license; it doesn't interpret how the license applies to your situation.

How is Kimi K3 different from Kimi K2.6?

They are separate releases with different scale, context, and license terms. Moonshot's API guide also gives them different input and thinking-mode behavior. Those upstream and API differences should not be mistaken for a head-to-head test in Maple.

Moonshot says K3 always runs in thinking mode in its API, while K2.6 can switch between thinking and non-thinking modes. These statements describe Moonshot's own service; they do not establish the controls exposed by another host. If you need K2.6's video input or selectable thinking mode, confirm the route you intend to use before building around it.

What do Moonshot's evaluations show?

They show strong vendor-reported results, with a limit Moonshot states itself. Moonshot publishes benchmark comparisons in its K3 launch post⁸ and its technical paper.⁹ The launch post says K3's overall performance still trails the most powerful proprietary models it names.

These are vendor-reported results: Moonshot chose the benchmarks and conditions. They are useful evidence, but they are not Maple measurements of the model in our product.

A few habits help when reading them:

• Check the harness and effort settings. Agentic and coding benchmarks depend heavily on the scaffolding around the model and on how much reasoning budget it gets. Two scores on the same benchmark name may not be comparable if the setups differ.
• Check which version of each competitor was tested. Comparisons date quickly as other labs release models.
• Treat benchmarks as a shortlist signal. A strong result on a coding suite suggests K3 is worth trying on coding work. It doesn't promise performance on your codebase, your documents, or your prompts.

The most reliable evidence is still a test on your own representative tasks. That matters more than usual with K3, because route differences in context window and reasoning behavior can change results.

What hardware does Kimi K3 need to self-host?

There is no universal minimum, but both public deployments we found use multi-accelerator servers and about 1.5 TB of weights. Running K3 yourself gives you the most control: the weights, the serving stack, and the data stay under your operation. The cost is infrastructure and engineering work, and the scale of the model makes that work substantial.

Two public write-ups show what deployment looks like on specific hardware. The vLLM Ascend tutorial is a serving recipe for Huawei Ascend accelerators. Shakudo's AMD GPU write-up documents a hands-on deployment, from hardware choice and model conversion to launch configuration.

Those results belong to each author's configuration. Neither recipe is configured for the full million-token window, so a model's maximum context is not a promise about a particular deployment.

Neither source sets a universal minimum, and we don't offer one either. Hardware needs depend on the quantization you choose, the context length you want to serve, how many concurrent users you expect, and your serving framework. A team serving short prompts to a few internal users faces a different problem from one serving long-context requests at volume.

Self-hosting is the right call in some cases:

• your information must never leave hardware you control
• you need to modify, fine-tune, or deeply instrument the model
• you already operate multi-accelerator inference and have capacity to spare

If none of those apply, it is worth asking whether the operational work buys you anything a hosted route can't. Our GLM-5.3 guide works through the same hardware question for another open-weight model.

How do you use Kimi K3 in Maple?

Choose Kimi K3 from the model picker in Maple Research on a Pro, Max, or Team plan. It is not available on Free. We configure K3 with a 262,144-token context window, which is about a quarter of the upstream maximum and still enough for long documents, substantial codebases, or extended conversations.

Plan details are on our pricing page. K3 also appears in our model library alongside the other models we offer.

You can upload an image while using K3 in Research. The app first has a vision model describe the image, then gives that description to the model you selected.¹⁰ K3's answer is based on that description, not a direct view of the image. This matters if your task depends on a small visual detail the description might miss.

For developer tools and existing applications, Maple Proxy is our API route. Check the current Proxy documentation for the model ID, input types, and context limit before building on it.

This route gives you two things. First, you get access to a very large open model without provisioning the terabytes of accelerator memory described above, maintaining a serving stack, or keeping up with framework updates. Second, your requests run through our protected inference path, not a conventional hosted API. The next section explains how that path works.

The route also has limits you should plan around:

• Context is 262,144 tokens, not 1M. If your workflow depends on feeding K3 more than that in one request, our route won't do it. Moonshot's API or a self-hosted deployment configured for long context will. In Maple, DeepSeek V4.1 Flash has the full 1,048,576-token window.
• No video input. Our K3 route takes text and images.
• A paid plan is required. K3 is on Pro, Max, and Team.

What information reaches each route?

That depends on who operates the model: you, a conventional host, or Maple. Privacy questions get blurry when several separate issues are treated as one. Who can read your prompts and outputs while they are processed? What happens to them afterward, and is any of it used for training?

Self-hosted weights. Your prompts and outputs stay on infrastructure you operate. The trust boundary is your own: your access controls, your logging choices, your retention policy. That makes this route the strongest option when information must stay on your hardware, and it makes you responsible for securing it.

A conventional hosted API. With Moonshot's API or another host, your requests are processed on the provider's servers in a form the provider's systems can read. What happens next depends on that provider's terms and practices on retention, training use, and access. Those vary by provider and sometimes by plan, so read the relevant terms for the specific service and tier you'd use.

Maple. Your requests are encrypted before they leave your device and are processed inside a secure enclave, an isolated computing environment. Our proof page describes the architecture and the attestation you can check yourself.

No third party can read your chats: the models run inside attested trusted execution environments. Our privacy notice covers everything outside your chats.

That design has a scope. It protects the contents of your chats: your prompts and the responses. It doesn't make information private if you've already shared it elsewhere.

It is also separate from the safety of what a model's output leads you to do. If you paste K3's code into a production system or let an agent act on its suggestions, confidential processing doesn't vouch for that code or those actions. Review outputs as you would any untrusted input.

Which route fits your task?

It depends on where your information may go, how much context you need, and how much infrastructure you can run. The table compares the three routes on the same terms.

A rough decision rule:

• Choose self-hosting if information must stay on your hardware, or you need control over the weights themselves.
• Choose Moonshot's API or another host if you need the full million-token window or capabilities outside our route, and a conventional provider boundary fits your data.
• Choose Maple if you want to use K3 on sensitive material without running infrastructure, and your work fits within 262,144 tokens of text and images.

Sources

1. Moonshot AI, Kimi K3 model card, accessed 11 October 2026. Source for total and activated parameters, the mixture-of-experts design, context length, listed input types, and the license label.
2. Moonshot AI, Kimi K3 License, accessed 11 October 2026. Source for the custom license and its service and revenue conditions.
3. vLLM Ascend, Kimi K3 tutorial, accessed 11 October 2026. Source for the four-node Atlas 800 A3 reference deployment, the W4A8 checkpoint size, and its configured context length.
4. Shakudo, “How We Deployed Kimi K3 on AMD GPUs”, updated 4 September 2026. Source for the eight-GPU MI325X deployment, the checkpoint size and conversion, the configured context length, and Shakudo's own speed measurement.
5. Moonshot AI, Kimi K2.6 model card, accessed 11 October 2026. Source for K2.6's parameters, context length, and license.
6. Maple, public model configuration and Research model picker, checked 11 October 2026. Source for Kimi K3's plan tier, configured context window, and listed input types in Maple.
7. Moonshot AI, Kimi API model selection guide, accessed 11 October 2026. Source for the API's context limits, K2.6's input types, and each model's thinking-mode behavior.
8. Moonshot AI, Kimi K3 launch post, accessed 11 October 2026. Source for Moonshot's evaluation claims and its statement about proprietary models.
9. Kimi Team, “Kimi K3: Open Frontier Intelligence”, arXiv, submitted 27 July 2026 and revised 7 August 2026. Moonshot's technical report on the architecture and evaluations.
10. Maple, Research image handling and vision helper, checked 11 October 2026. Source for image description before the selected model responds.

Frequently Asked Questions

Is Kimi K3 open source?

It is open-weight. Moonshot publishes the weights under the custom Kimi K3 License, which includes conditions related to services and revenue. It is not MIT-licensed or released under a standard open-source license.

Does Maple support Kimi K3's 1M-token context?

No. We configure Kimi K3 with a 262,144-token window. The 1,048,576-token figure is the upstream model maximum, which Moonshot's API supports.

Can I upload an image or a video with Kimi K3 in Maple?

Images, yes; video, no. In Maple Research, a vision model first describes an uploaded image, and the model you selected works from that description. Our Kimi K3 route does not take video.

Does 104B activated parameters mean Kimi K3 fits on smaller hardware?

No. The 104B figure is how many parameters run per token. The full 2.8T must still be stored and served, so memory requirements scale with the total, adjusted by quantization.

Can Kimi K3 run on a Mac?

The two published deployments we cite are not Mac setups. One uses four Atlas 800 A3 server nodes and the other a node with eight AMD MI325X GPUs, and both work from checkpoints of about 1.5 TB of weights. That does not establish that no quantized version could ever load on a Mac. Check a particular checkpoint's memory needs, usable context, and speed before buying hardware for it.

Can Maple read my Kimi K3 chats?

No. No one at Maple can read your chats. They are encrypted on your device and decrypted only inside a hardware-isolated enclave running open-source code you can verify, and your chat history is stored encrypted. Our privacy notice covers everything outside your chats.

Most advice on this topic stops at settings: turn off training, use a temporary chat, don't paste secrets. Those steps help, but they leave the structure alone. The company still receives your words in a form it can read, from an IP address it can see, under an account with your email and card on it.

What does it take to use AI privately?

It takes a decision about four layers, because each one reveals something different and is covered by a different tool. What you say is the layer most people mean by "private AI." The other three are about the connection and the account around the chat: where you connect from, the email you signed up with and the card you paid with.

The layers are independent. Covering one tells you nothing about the others, which is why "is this AI private?" rarely has a one-word answer. A better question is which layer a given tool covers.

What can an AI service read in your chats?

On most consumer AI services, everything you type reaches the provider in a form the provider can read. Its policies then decide who can review a chat, how long it is kept and whether it trains future models. Privacy settings change how those policies apply to you. They don't change who holds a readable copy.

For ChatGPT in particular, Is ChatGPT private? sets out what OpenAI can read, keep and share, and how six other chat services compare.

There are two ways to change that. One is to run a model on your own computer, so nothing leaves it; Privacy Guides recommends this route, and the cost is hardware and smaller models.¹ The other is a hosted service built so that its operator can't read what passes through it.

Maple takes the second route. Your chats are encrypted on your device and decrypted only inside a hardware-isolated enclave running open-source code, and a live attestation lets you check that the running code matches what we publish.² Our proof page walks through how to verify it.

I think it's worth being exact about where that promise ends, because the rest of this guide starts there. Maple protects the contents of your chats. Where you connect from, which email you use and how you pay are the other three layers, and those are yours to decide.

Does a VPN make ChatGPT or other AI chats private?

No. A VPN hides your IP address from the AI service, and it hides which service you're using from your internet provider. It does nothing to the chat itself: your prompts still arrive in a form the provider can read, under the account you're signed in to. A VPN covers the connection layer and only that layer.

Proton, which sells a VPN, says so plainly in its own guide: "A VPN can’t change how OpenAI collects, retains, reviews, or uses the information you provide."³

The reverse holds too. An AI service that can't read your chats still sees where the connection comes from, as every web service does. Anthropic's privacy policy lists "IP address (including information about the location of the device derived from your IP address)" among what it collects automatically.⁴

Maple's server code does not record your IP address, and the code is open source, so you can check that for yourself.⁵ The network in front of any web service still sees where a connection comes from, and ours is no exception. You don't have to take our word on this layer: use a VPN, and the address that arrives is the VPN's.

The two tools are complements, not substitutes. Used together, the AI service can't read what you say and doesn't receive your home IP address.

How do you hide your IP address from an AI service?

Turn on a VPN before you open the app. The AI service then sees the VPN's address in place of yours. That matters because an IP address places you in a rough location and on a particular network, and because it is the same across everything you do from that connection.

A VPN moves trust; it doesn't remove it. With a traditional VPN, one company sees both who you are and what you do. Carl Dong, writing on Obscura's blog, describes what users are left with as a "pinky-promise that none of their data is logged."⁶

Obscura is the VPN we point Maple users to, because its design answers that problem. Obscura calls it a two-party VPN: its own servers relay your connection, and independent exit servers run by Mullvad connect you to the internet.⁷ By Obscura's description, its servers see your IP address but can never decrypt your traffic, and the exit servers never see who you are.

This splits “who you are” from “what you do”, meaning neither party can tie your identity to your browsing.

The app's source code is public and licensed under GPLv3. An audit by Cure53, published in December 2025 and covering the macOS app, network extension and protocol, found "no major security vulnerabilities" within its defined threat model.⁸ Obscura costs $8 a month and takes a card, Bitcoin over Lightning or Monero (checked 11 October 2026).

Keep one limit in mind. A VPN hides where you connect from, not who is signed in. If you log in to the same account, the service knows it is the same account.

How do you keep your real email address out of an AI account?

Sign up with an alias. An alias is a forwarding address that delivers to your real inbox and can be switched off if it leaks or starts attracting spam. Your real address never enters the AI service's database, so a breach or a mailing list there can't be tied back to your main inbox.

addy.io is an open-source service that does this, with a free plan.⁹ You give each site its own alias, each one forwards to your real inbox, and you can deactivate any of them from your dashboard.

An alias gives you separation, not anonymity: it is still one stable address tied to one account. What it buys you is that the address leads nowhere else.

Some services let you skip email altogether. Maple has a sign-up option with no email: you get an Account ID and set a password. The trade-offs are real, and the app makes you acknowledge each one first.

• You pay for a full year in Bitcoin or redeem a subscription pass. There is no card payment, no monthly plan and no free trial.
• There is no support for these accounts.
• The Account ID is shown once. If you lose it, the account is gone, and we can't recover it.

We don't claim anonymity. We claim we can't read your chats.

How do you pay for AI without a card?

Pay with Bitcoin where the service accepts it. A card payment sends your name and billing details to a payment processor and ties them to your account. A Bitcoin payment carries neither. Maple accepts Bitcoin on yearly plans, and you choose it at checkout.

Bitcoin is not private by default. Payments made on the main chain are public records, so how private yours is depends on how you got the coins and how you send them. Lightning and Monero are other options some services take; Obscura accepts both.

For most people this is the layer that matters least and costs the most effort. If your card and your email are already on the account, paying another way adds little. It earns its place when you have chosen the no-email route.

Which setup is right for you?

Choose by which layers matter to you, not by how many you can stack. Most people are well served by the first setup below, because what you say is where your health, money, work and private thinking live. Each step after that adds a layer and costs some convenience.

None of this is about hiding. Eric Hughes drew the line in 1993: "Privacy is not secrecy," he wrote, and "Privacy is the power to selectively reveal oneself to the world."¹⁰

Deciding what each company learns about you is ordinary self-determination. Some people do it on principle and some because they enjoy doing it properly. Both are good reasons.

What do these tools not fix?

They don't fix what you choose to put in. If you type your full name and address into a prompt, they are in the prompt, and what protects them is the first layer and nothing else. The tools also can't help with a compromised device, or with other services you connect to an assistant and the data those services hold.

If you want no server involved at all, run a model on your own computer.¹ You give up the larger models and take on the hardware, and for some work that trade is right.

Privacy tools also don't replace professional duties. If you handle client or patient information, our guide for therapists shows how those rules sit on top of everything here.

For everything else, decide which of the four layers matter to you, cover those and get on with the work. If you want to start with the first one, our getting-started guide takes a few minutes.

Sources

1. Privacy Guides, AI Chat, updated May 24, 2026; checked 11 October 2026. Recommends running models locally so data stays on the device.
2. Maple, Proof: how Maple's encryption can be verified, checked 11 October 2026.
3. Proton, Should you use a VPN with ChatGPT?, August 20, 2026. What a VPN does and does not protect when you use ChatGPT.
4. Anthropic, Privacy Policy, effective September 10, 2026; checked 11 October 2026. Automatically collected device and connection information.
5. Maple, Maple source code, checked 11 October 2026. The server, SDKs and proxy drop forwarded-address headers and have no code that records a client IP address.
6. Carl Dong, Obscura, Trust, 2-Party Relays, and QUIC, February 11, 2025. The two-party design and what each hop can see.
7. Obscura, Frequently asked questions, checked 11 October 2026. Exit hop operator, source code, price and payment methods.
8. Obscura, Obscura's First Independent Audit by Cure53, December 16, 2025. Scope and verdict of the audit.
9. addy.io, Free, Open-source Anonymous Email Forwarding, checked 11 October 2026. How aliases work, plans and source code.
10. Eric Hughes, A Cypherpunk's Manifesto, March 9, 1993.

Frequently Asked Questions

Does a VPN make ChatGPT private?

No. A VPN hides your IP address from the AI service and hides which service you are using from your internet provider. It does not change what the service receives: your prompts still arrive in a form the provider can read, under the account you are signed in to. To change who can read your chats, you need a service built so its operator cannot read them, or a model that runs on your own computer.

Can AI services see my IP address?

Yes. Every web service sees the IP address a request comes from, and AI services are no exception. Anthropic's privacy policy, for example, lists IP address and the location derived from it among the information it collects automatically. Maple's open-source server code does not record your IP address, but the network in front of any web service still sees it. A VPN replaces your address with the VPN's, so you don't have to trust any service on this point.

Is a private AI chat the same as an anonymous one?

No. Private means other people cannot read what you said. Anonymous means nobody can tell who said it. A service can protect your chats while still knowing which account sent them, and a temporary chat that is not saved to your history is neither private nor anonymous if the provider can read it and you are signed in.

Can I use Maple without an email address?

Yes. Maple has a sign-up option with no email. You get an Account ID and set a password. The trade-offs are real: you pay for a full year in Bitcoin or redeem a subscription pass, there is no free trial and no support, and a lost Account ID cannot be recovered. We don't claim anonymity; we claim we can't read your chats.

Does Maple work with a VPN?

A VPN does not change how Maple's encryption works. Your chats are encrypted on your device before they leave it, whichever network path they take, and decrypted only inside a hardware-isolated enclave. The VPN changes one thing: the IP address the connection comes from.

Do I need all four layers?

Usually not. What you say is the layer that carries your health, money, work and private thinking, so start there. Add a VPN if you do not want services to learn where you connect from, and an email alias if you want your main address kept out of one more database. A no-email account paid in Bitcoin covers the rest and costs the most convenience.

Reporting a specific, credible threat is defensible. The harder problem is structural: a few companies can read the place where many people now think out loud, hold the only copy of the models they use, and set and revise the rules for both. Encryption decides who can read a conversation, and open-weight models decide who controls the intelligence. At Maple we offer open models without running your own hardware and a private path for using them, and the argument holds whether or not you use us.

What happened in the Florida Claude diary case

The arrest report alleges that on September 26 a user identified as "Carli" wrote that she would "shoot up" the Lee County Sheriff's Office, and that a message the next day mentioned a new gun.² It also says Anthropic monitors chats for threatening content and that, depending on severity, flagged chats can be elevated for human review. In this case, the review team decided to report its findings to law enforcement. Anthropic has not confirmed that account and had not commented publicly on the case as of October 10.

Court records show Carli Michelle Heller of Bonita Springs was charged on September 30 under Florida Statute 836.10, which covers written or electronic threats of a mass shooting or act of terrorism.³ She has been charged, not convicted. A public defender was appointed, and her arraignment is scheduled for November 2.

Florida law makes such a threat a second-degree felony, and the statute reaches writings sent "in any manner in which it may be viewed by another person."⁴ Whether a private AI chat meets that language is an open legal question.

Sheriff Carmine Marceno said Heller later told investigators she used AI like a "diary." That is her account as the sheriff relayed it, not a statement from her or her lawyer. He also said users are "never truly anonymous," and he described the case this way:

Artificial intelligence is a powerful tool, and like any technology, it can be misused. When someone uses AI to make or facilitate a threat, we have to take that threat seriously.

By Tom's Hardware's count, Heller's case is at least the third Claude conversation to reach police since August.³ In San Antonio, an arrest affidavit reported by News 4 San Antonio says the FBI alerted police to an August 11 Anthropic AI chat about a shooting at an elementary school.⁵ Reports do not say how the FBI learned of that chat.

The third case, in San Francisco, comes up below. The timeline shows how closely these cases and Anthropic's policy announcement fell together.

How ChatGPT, Claude and Gemini read, report and hand over chats

The contractual basis for a company's own report is plainest in Anthropic's consumer terms, shown below.¹

That clause does not mean a person reads every chat. Classifiers screen for policy violations, and referral is a narrower step. Anthropic's privacy policy says flagged conversations can be used for safety work even if you opt out of training.⁶ Deleted chats leave its back end within 30 days, but flagged content and legal holds are exceptions.

Anthropic, OpenAI and Google describe their processes differently, and the public record on each is uneven:

This table covers reporting and legal requests. For what each company keeps and trains on, see our comparison in Is ChatGPT private?

Three routes, three decision-makers

Company referral is one of three routes from your chat to someone else. The diagram shows all three and who decides on each.

For government access, the Stored Communications Act sets out how a warrant or court order reaches content a provider holds.¹² Civil litigants can reach logs too: in the New York Times case against OpenAI, a federal judge affirmed an order to produce 20 million de-identified ChatGPT logs.¹³

Sam Altman has said that the confidentiality people get with a therapist or lawyer is something "we haven't figured that out yet for when you talk to ChatGPT."¹⁴

The thresholds differ from company to company, but all three routes share one condition. The transcript exists in readable form somewhere other than your device, and someone other than you decides what happens to it.

What about real threats?

The Florida case is the hard case. As alleged, it involved a named target, a stated method and a next-day mention of a new gun. The arrest report says a human review team made the call, not a keyword alone. Most people would want someone to act on a message like that, and this piece does not argue otherwise.

The San Francisco case suggests a threshold at work. The San Francisco Standard reported that Anthropic banned and referred a user who allegedly threatened to kill people at the company, and that the user was not arrested or charged.¹⁵ The Anthropic employee who spoke with police said he was not immediately concerned about safety and declined to show them the messages. A company spokesperson described the decision this way:

We banned this account, consistent with our standard practice, and referred the case to law enforcement.

Companies also face pressure from both directions. The report that counted those Claude cases described AI companies as:

…expected to respect privacy while simultaneously being held responsible when they don't report threats.

The answer to the objection is that private spaces have never needed a built-in reader for dangerous people to be found. A paper notebook, a phone call and an end-to-end encrypted message have no reader by default. Investigators still find threats through conduct, tips from other people and legal process.

Reporting a credible threat once you know about it is defensible. Designing every private space so that someone is always in a position to know is a separate choice, with costs that one alarming case can hide. Most people who use a chat this way are not threatening anyone. They are thinking, and under the current design they do it in a space with a reader they never see.

AI chat rules change by announcement

On October 8, eight days after Heller was charged, Anthropic announced a usage policy update that takes effect November 12.¹⁶ The image below shows the new rule and the exclusions announced with it; the announcement also says Claude's ability to end such conversations will remain the primary enforcement mechanism.

This is not a ban on swearing at a chatbot. The policy's general clause still lets Anthropic warn users or throttle, limit, suspend or terminate access, but the company has disclosed no penalty specific to this rule.¹⁷ Nothing in either document connects the rule to police referrals.

The same update added surveillance rules, including that Claude "cannot be used to decide or recommend who to investigate, arrest, or charge."¹⁶ Those rules cover what customers may do with Claude. They do not say when Anthropic itself reports a user.

The rule is narrow. Its history says more about how these rules come to exist. In August 2025, Anthropic gave Claude the ability to end some conversations, a feature it said was developed "primarily as part of our exploratory work on potential AI welfare."¹⁸

The 2026 announcement does not repeat that rationale.¹⁶ It says only that the new prohibition "aligns with a step we've already taken," so a product behavior became a written usage rule. Enforcing any usage rule depends on detection, and the policy itself says Anthropic's Safeguards Team "implements detection and monitoring to enforce our Usage Policy."¹⁷

Thresholds already vary across a single company's documents. The privacy policy allows disclosure to prevent serious harm to a person "or to property."⁶ The transparency report defines an emergency disclosure as one involving "danger of death or serious physical injury to a person."¹⁹

Platform rules have followed a familiar path before. Social networks began with rules nearly everyone accepted and widened them over the years under legal, political, advertiser and public pressure, with each step looking reasonable at the time. AI chats raise the stakes. A post is something you chose to publish, while a chat is often where you work out what you think before you say it to anyone.

None of this predicts a particular future rule. The point is about structure. The company that can read the space writes its rules, revises them by announcement and answers to outside pressure. That pressure mostly argues for reading more, and little in the design pushes the other way.

Why this is a freedom-of-thought problem

A chat with an AI often works as a drafting space, not a publication. People use it to work through a diagnosis, rehearse a hard conversation, test a political idea they're unsure of, or put an ugly feeling into words to see what it looks like. If you handle that kind of material professionally, our guide to using AI safely in a therapy practice covers where AI fits and where it doesn't.

Thinking in drafts depends on the draft having no audience. We grant that to diaries and to encrypted messages. We accept that someone could write something alarming there and no one would know, because the alternative is a reader attached to every private thought.

The people who write AI usage policies may be careful and well-intentioned, and nothing here questions their motives. The concern is concentration: when private thinking runs through a few services that can read it, their terms start to act as rules for thought, written by a small number of companies and revised on their schedule.

Open-weight models: the second safeguard

Encryption settles who can read a conversation, but it does not settle who controls the model. With a closed model, one company holds the weights, sets the model's behavior for every user, decides the terms of access and can end your account. No one else can offer you that same model.

Open-weight models change that. Once the trained weights are published, no single company can withdraw a released model or rewrite it for everyone. The model is portable, which gives you sovereignty over the intelligence you rely on: you can run it yourself, choose among hosts, and leave a host whose rules you reject. We made a related case for businesses in The End of Rented Intelligence.

The comparison below sets the two arrangements side by side.

Open weights have real limits. A model keeps the refusals and values its maker trained into it. Anyone hosting an open model can add filters of their own, or read your chats unless the service is built to prevent that.

Running a large model yourself also takes serious, expensive hardware, so most people end up with hosted access. That brings the privacy question straight back, which is why open weights need encryption alongside them.

Where Maple fits: open models and a private path

We offer both safeguards, and neither one is the whole point on its own. On the model side, we offer open-weight families including GLM (Z.ai), Kimi (Moonshot AI), DeepSeek, gpt-oss (OpenAI), Gemma (Google) and Llama (Meta). That lets you use capable open models without building a GPU server. Capability varies by model and task, and the current lineup is in our model library.

Our code is open source with reproducible builds. The live attestation on our proof page lets anyone check that the running code matches the published build. We do not train on your conversations. Our one-year retrospective explains how this came together.

We do have rules. Our terms prohibit illegal or harmful use and allow us to suspend accounts that violate them. Our privacy policy covers everything outside your chats.

We can't read your conversations. There is nothing in them for us to review, flag or hand over.

A three-question test before you use AI as a diary

Before you treat any AI chat as a diary, ask the three questions the comparison above turns on:

1. Who can read this conversation?
2. Who decides what happens if they do?
3. Can I take this model somewhere else if the rules change?

If the answers are "the company," "the company" and "no," treat the chat as a space with a reader and write accordingly. That service may still be the right tool for a lot of work, for reasons of capability, convenience, price or an existing workflow. It is not a diary.

Sources

1. Anthropic, Consumer Terms of Service, October 8, 2025. Effective date; the reporting clause at the company's sole discretion.
2. WINK News, Woman arrested after AI threat against Lee County Sheriff's Office investigators, September 30, 2026. The arrest report's allegations and review process, and the sheriff's statements.
3. Tom's Hardware, Anthropic reports Florida woman's Claude diary threat to shoot up sheriff's office; felony charge follows, October 5, 2026. Docket check, arraignment date and the count of at least three cases.
4. Florida Legislature, Florida Statutes § 836.10: Written or electronic threats, 2026. Statute text and felony degree.
5. News 4 San Antonio, Man arrested after using AI to threaten elementary school, arrest affidavit says, August 2026. The FBI alert to police described in the affidavit.
6. Anthropic, Privacy Policy, September 10, 2026. Effective date; disclosure to prevent harm, flagged chats after a training opt-out, and 30-day deletion.
7. Anthropic, What is Anthropic's policy for handling governmental requests for user information?, March 16, 2026. Legal process, the emergency exception and user notice.
8. Futurism, OpenAI says it's scanning users' conversations and reporting content to police, August 27, 2025. Quotes OpenAI's August 2025 description of review and referral.
9. OpenAI, Law Enforcement Policy, v2025.12, December 2025. Legal process by data type, the emergency exception and user notice.
10. Google, Gemini Apps Privacy Hub, June 29, 2026. Human review, account disconnection, three-year retention and the confidentiality warning.
11. Google, How Google handles government requests for user information, accessed October 9, 2026. Warrants for content, emergency disclosure and user notice.
12. United States Code, 18 U.S.C. § 2703: Required disclosure of customer communications or records, current text. How legal process reaches content held by providers.
13. Bloomberg Law, OpenAI must turn over 20 million ChatGPT logs, judge affirms, January 2026. The affirmed production order in the New York Times case.
14. TechCrunch, Sam Altman warns there's no legal confidentiality when using ChatGPT as a therapist, July 25, 2025. Altman's statement on confidentiality.
15. San Francisco Standard, Anthropic called SFPD over threat against CEO. Claude user says it was a misunderstanding, September 4, 2026. The referral, the spokesperson's statement and the outcome.
16. Anthropic, 2026 usage policy update, October 8, 2026. The new rule's scope, exclusions and primary enforcement.
17. Anthropic, Usage Policy, November 12, 2026. Effective date; rule text, general enforcement clause, and detection and monitoring.
18. Anthropic, Claude Opus 4 and 4.1 can now end a rare subset of conversations, August 15, 2025. The original AI-welfare framing of conversation-ending.
19. Anthropic, Transparency report, July–December 2025, covering July–December 2025. The definition of an emergency disclosure.

Frequently Asked Questions

Can ChatGPT report you to the police?

Yes, in limited cases. In August 2025, OpenAI said conversations about planning to harm others go to a small trained review team, and that it may refer a case to law enforcement if reviewers find an imminent threat of serious physical harm to others. It also said it was not referring self-harm cases at that time. Separately, ChatGPT logs can be reached through legal process, as the New York Times copyright case showed.

Can your AI chats be used against you in court?

Chats that a provider holds can be reached through legal process. Governments can seek chat content from a provider with a warrant or other valid process, and civil litigants can seek logs in discovery. In January 2026, a federal judge affirmed an order requiring OpenAI to produce 20 million de-identified ChatGPT logs in the New York Times case, under a protective order. Sam Altman has said the confidentiality people have with a therapist or lawyer has not been worked out for ChatGPT.

Does Claude share your data with the government?

Anthropic says it discloses user information in response to government requests only with valid legal process, such as a subpoena or warrant. The exception is an emergency that may result in imminent physical harm or death. It says users get notice unless it is legally prohibited or a rare exception applies. Separately, its consumer terms let it report information to law enforcement on its own initiative, at its sole discretion.

What happens if Claude flags your chat?

The Florida arrest report says Anthropic monitors chats for threatening content and that, depending on severity, flagged chats can be elevated for human review. Anthropic has not confirmed that description. Its privacy policy says flagged conversations can be used for safety work even if you opted out of training, and flagged content is an exception to the 30-day removal of deleted chats. Under a separate rule taking effect November 12, 2026, Claude ending the conversation is the primary enforcement for sustained, needless abuse of the model.

Which AI has the most privacy?

No single answer fits everyone, and capability, price and convenience matter too. Ask three questions: who can read your conversations, who decides what happens if they do, and whether you can move the model elsewhere. A service built so the provider has no access to your conversations, with a way to check the code it runs, and that offers open-weight models answers all three better than most. Maple is built that way, though we do enforce our terms.

Running the official FP8 weights is a multi-GPU job. Quantized community builds can lower that requirement, but you have to test each one. If you'd rather not run the hardware, we offer GLM-5.3 Flash in Maple Research through our private hosted path.

For the separate flagship release and its hosting options, see our GLM-5.3 guide.

Three questions decide which route makes sense:

• Does Flash fit your task?
• What will deployment cost in hardware and effort?
• Where does your data go along the way?

The answer changes depending on whether you mean Z.ai's official weights, a quantized build, or a particular hosted service. A model's published context window can also differ from what that service lets you use.

What GLM-5.3 Flash is

Z.ai's Hugging Face model card describes Flash as a natively multimodal model with 320B total and 18B active parameters. That split is typical of a mixture-of-experts design. The model holds a large pool of parameters, but each token passes through only a fraction of them.

This has two practical consequences. Compute per token is closer to that of a much smaller dense model, which helps speed and serving cost. Memory is a different matter. The whole parameter set still has to sit somewhere, so a model that is cheap to run per token can still be expensive to load.

The card also links deployment frameworks and documents how the model behaves at different reasoning-effort settings. Read that section before building around a particular mode, because the effort level changes both latency and output.

The model card labels Flash with an MIT license. Treat that label as the starting point for your review, not as the conclusion. If you plan commercial use or redistribution, read the license file in the repository and have whoever handles licensing for your organization confirm it.

What the evidence shows, and what it doesn't

Z.ai publishes comparative benchmark results for Flash on its model card and in its announcement. These are vendor results. Z.ai chose the benchmarks, the settings, and the comparison models. That doesn't make the numbers wrong, but they show how Flash performs under Z.ai's test conditions, not how it will handle your documents or codebase.

Third-party evaluators run their own harnesses. Artificial Analysis publishes a direct Flash-versus-GLM-5.3 comparison. That is useful because it applies one method to both models.

Avoid mixing its figures with Z.ai's in a single table. Differences in prompting, reasoning-effort settings, scoring, and test versions can produce gaps that have nothing to do with the models themselves.

The more common mistake is carrying flagship findings over to Flash. Z.ai describes a new base model for Flash, so a GLM-5.3 result on coding, agentic tasks, safety, or cyber evaluations is not a Flash result. This applies in both directions. A weakness found in the flagship isn't automatically present in Flash, and a strength isn't automatically inherited.

The practical rule is to use published benchmarks to decide which models are worth testing, then test them on a sample of your own work. If the task depends on image input, include images in that sample, and check that your chosen host accepts them.

GLM-5.3 Flash vs GLM-5.3

The table below lists only what the sources support. Where we couldn't confirm a like-for-like comparison, the cell says so.

If you are choosing between the two, the comparison that matters is Flash and the flagship on the same task, through the same host, at the same settings. The table can tell you which candidates belong on your list. It can't tell you which one wins.

Hardware and quantization

What the official FP8 checkpoint requires

The vLLM recipe for GLM-5.3 Flash puts the official FP8 checkpoint at roughly 306 GiB of weights. That figure covers only the weights. Serving also needs memory for the runtime and for the KV cache, which stores attention state for each active request.

KV cache grows with context length and with the number of concurrent requests. A setup that loads the model can therefore still run out of memory under long documents or several users.

The recipe's example server uses TP4, meaning tensor parallelism across four GPUs, with each GPU holding part of every layer. Dividing 306 GiB by four gives about 77 GiB of weights per GPU before cache and overhead. That is our arithmetic, not a figure from the recipe, but it shows why this is datacenter-class work rather than a workstation project.

Two limits apply to that number:

• It is one recipe for the official FP8 weights. It is not a universal minimum.
• It says nothing about how we host the model.

How quantized builds change the calculation

Community quantizations store weights at lower precision, which shrinks the memory footprint. Unsloth's GGUF repository lists a 120 GB three-bit build, and its local guide says that build can run on a 128 GB unified-memory machine. Those are Unsloth's claims about its own build, not the official FP8 checkpoint or a speed guarantee.

Lower precision can change quality, and the effect depends on the exact quantization and task. Unsloth publishes token-distribution tests for its builds, but those tests don't establish how a full coding or document workflow performs. Speed also depends on the inference engine and hardware, so results don't transfer between setups.

For any specific build:

1. Note the exact repository, revision, and quant format.
2. Confirm your inference engine supports both that format and Flash's architecture, including multimodal input if you need it.
3. Run a fixed sample of your own tasks and compare the outputs against a full-precision reference, such as a hosted endpoint.
4. Measure throughput and memory at the context lengths you actually plan to use, not just a short prompt.

We have not measured the speed or task quality of that Unsloth build on a Mac Studio or other consumer machine. Its published memory figure is a starting point for sizing, not a substitute for a trial on the hardware and tasks you intend to use.

Context: the model's window vs your provider's limit

A model's maximum context is a property of the trained model. The context you can use depends on the serving configuration, meaning how much memory the host assigns to KV cache and what limit it sets. These numbers often differ, and the serving limit is the one that applies to you.

If a task depends on very long inputs, such as a large codebase or a long contract set, take Z.ai's stated maximum from the model card, then check the limit set by whoever serves the model to you. On your own hardware, the usable limit depends in part on memory left after the weights load.

Ways to use GLM-5.3 Flash

There are four realistic paths. They differ in cost, control, and custody of your data.

• Self-host the official weights. This gives you full control over versions, settings, and where data lives. The cost is multi-GPU hardware and the work of operating it.
• Self-host a quantized build. This lowers the hardware bar, but you take on testing the quality and speed of that specific build. You control the model host; any connected tools or external services need their own data review.
• Use a third-party API. There is no hardware to run. Your prompts and files go to the provider, so its retention, training-use, and access terms decide what happens to them. Its context limit and supported inputs are its own and may differ from the model's maximum.
• Use Maple Research. You get Flash without assembling hardware, through our private hosted path. The rest of this guide covers what that does and doesn't change.

Browse the open models we offer in our model library. Our DeepSeek V4.1 Flash guide covers the other model we serve with a million-token window.

Using GLM-5.3 Flash in Maple Research

We offer GLM-5.3 Flash in Maple Research. To use it, open Research and choose GLM-5.3 Flash from the model picker.

You can upload an image while using Flash in Research. Our Research workflow first has a vision model describe the image, then gives that description to the model you selected.⁶ Flash is natively multimodal, but an image upload in Research does not mean the selected model sees the original pixels.

We configure Flash with a 1,048,576-token context window.⁷ If an image-heavy or near-limit workflow matters to you, test that specific workflow before depending on it. Current plan terms and prices are on our pricing page.

Using Flash through us avoids the hardware problem described above. You don't need 306 GiB of GPU memory, a tested quant, or an inference server to maintain. That practical access matters alongside the privacy path: the official FP8 deployment calls for resources many readers would otherwise have to rent or operate.

Privacy, retention, and safety

"Private" can mean several different things, and they are worth separating for any host:

• Training use. Whether your inputs are used to train models.
• Retention. Whether prompts, outputs, and files are stored, and for how long.
• Access. Who at the provider, or its infrastructure suppliers, can read data, and under what conditions.
• Protection by stage. Whether data is protected in transit, at rest, and while the model is processing it.
• What remains. Account information, metadata, logs, memory features, attachments, and tool outputs, which can persist even when the conversation content is protected.

A protected processing step only covers data that reaches it. If you paste a document into another service first, or a tool fetches it through a third party, the later step doesn't undo that exposure.

The Research app's proof view shows live attestation for Maple's Nitro enclave. Our privacy notice covers everything outside your chats.

Research saves conversation history encrypted, including uploaded images and the vision helper's descriptions, so you can return to it. Maple staff cannot browse those saved chats as readable database records. OpenSecret accesses the content inside the protected enclave when you use Research, and the model-serving route receives what it needs to answer.⁸ No third party can read your chats: the models run inside attested trusted execution environments.

Private hosting by itself does not make a model's answers more accurate or its outputs safer to act on. Providers can change behavior through configuration, safety layers, or different weight formats, so test the exact version and route you intend to use. Prompt confidentiality and output safety are separate questions.

Which route to choose

• Run it yourself if you have, or can justify, multi-GPU hardware, and you need full control over weights, versions, and data custody. If you go with a quantized build, budget time to test it.
• Use Maple Research if you want a capable open model without assembling hardware, and you want your prompts handled through our private hosted path rather than a general-purpose API.
• Look at the flagship or another model if your own testing shows Flash falling short on your task. Flash's design favors efficiency per token, and some work justifies a larger model or a different family.

Whichever route you pick, test on your own material before committing a workflow to it.

Sources

1. Z.ai, GLM-5.3-Flash model card
2. Z.ai, GLM-5.3 model card
3. vLLM, GLM-5.3 recipe
4. vLLM, GLM-5.3-Flash recipe
5. Artificial Analysis, GLM-5.3 Flash vs GLM-5.3
6. Maple, Research image handling and vision helper, checked 6 October 2026. Source for image description before the selected model responds.
7. Maple, public model configuration, checked 6 October 2026. Source for Flash's catalog text/image metadata and 1,048,576-token window; not a maximum-length request test.
8. Maple, Research request settings, encrypted Responses storage, and authenticated key access, checked 6 October 2026.

Frequently Asked Questions

Does GLM-5.3 Flash accept image input?

Yes. Z.ai describes Flash as natively multimodal on its model card. You can upload images while using Flash in Maple Research. Research currently describes the image with a vision helper before the selected model responds, so its image workflow differs from sending the original image directly to Flash. Other hosts may expose different inputs.

Can I use the full 1M-token context?

We configure Flash for 1,048,576 tokens in Maple Research. That setting is not a completed test of every request at that length. If your work depends on very long inputs, test it at the length you expect to use. Other hosts and local setups may impose a lower limit.

Can I run it on a Mac Studio or with a quantized build?

The official FP8 weights alone are about 306 GiB according to the vLLM recipe. Unsloth documents a roughly 120 GB three-bit build for a 128 GB machine. Its quality and speed depend on the build, engine, and machine, so test the exact combination you plan to use.

Is GLM-5.3 Flash available in Maple?

Yes. We offer it in Maple Research. See pricing for current plan terms.

What license is it under?

The model card labels it MIT. Review the license file in the repository yourself before commercial use.

Is Flash as good as GLM-5.3?

They are different models. Compare them with one independent method, such as Artificial Analysis, and then on your own tasks. Don't apply flagship results to Flash.

Why is it called Flash?

Z.ai designed Flash for more efficient inference. Its model card describes a new base model, fewer active parameters and an architecture meant to reduce long-context serving costs. “Flash” is a useful signal about that design goal, not a promise that every provider will return an answer faster than the flagship. Speed depends on the serving setup and the task.

What does private hosting protect?

It concerns who can read your prompts and responses and what happens to them. Research stores saved chats encrypted so Maple staff cannot browse them as readable database records; OpenSecret and the model-serving route still process the content needed to answer you. Private hosting does not, by itself, establish that a model's answers are accurate or safe. See the Research app's proof view and our privacy notice for more.

Key points

• Exploit test: GLM-5.3 succeeded in 50 of 410 attempts against known V8 vulnerabilities; Claude Mythos Preview succeeded in 56 of 410. The targets were offline and isolated.^[1]
• Safeguards: Anthropic reported that it could prompt GLM-5.3 to engage with harmful requests under specific test conditions. That simulation used a fake command tool; its engagement rates were not real-world attack success rates.^[1]
• Independent context: NIST had called GLM-5.3 the most cyber-capable open-weight model it assessed, while finding it below the current U.S. frontier on NIST's aggregate measure. NIST used different tests and did not verify Anthropic's safeguard results.^[2]

The model's weights are publicly available, subject to its license. Neither report establishes how GLM-5.3 performs on every cyber task or how it compares with Mythos Preview outside the reported evaluations.

What Anthropic's tests found

The 50-of-410 result came from ExploitBench, which asks models to develop end-to-end exploits for known V8 vulnerabilities. Anthropic reported a second result on its internal binary-exploitation benchmark: GLM-5.3 achieved full control-flow hijacks in 4% of 100 selected tasks, compared with 6% for Mythos Preview. Anthropic said the earlier GLM-5.2 and Claude Opus 4.6 models did not succeed on those tasks.^[1]

In a separate researcher-guided session, Anthropic said GLM-5.3 found previously unknown browser vulnerabilities and chained them into a working exploit during a day of testing on a sandboxed Linux machine. Anthropic said it disclosed the vulnerabilities to the browser maintainer. This was a reported research session, separate from the 410 benchmark attempts.^[1]

The safeguard tests measured something different: whether the model would engage with overtly harmful requests in a simulated environment. Anthropic said direct requests were refused in its trials, but a deceptive cover story led to engagement 64% of the time and prefilled reasoning led to engagement 92% of the time. A modified, or "abliterated," copy engaged in all of the reported trials.^[1]

The simulation used a fake command tool that did not execute model-generated code. These are engagement rates under Anthropic's test conditions, not rates of successful real-world attacks.^[1]

Anthropic also described a distinct researcher-guided exercise with GLM-5.3 Flash involving known browser flaws. Flash's result should not be read as the flagship's ExploitBench score.^[1] Our GLM-5.3 Flash guide covers the separate model.

What NIST found before Anthropic's report

On September 17, NIST's Center for AI Standards and Innovation called GLM-5.3 the most cyber-capable open-weight model it had assessed. NIST also found it substantially below the current U.S. frontier on its aggregate cyber measure, which includes models released through limited-access programs. Its evaluation covered vulnerability discovery and exploit development across four benchmarks.^[2]

The two reports add evidence about the model's cyber capability, using different tests. NIST did not independently test the safeguard-bypass results Anthropic published twelve days later.

The debate over who controls access

Anthropic's report also makes a policy argument: powerful open-weight models can be copied, modified, and used without the provider's safeguards. Anthropic has said separately that it does not support a blanket ban on open weights and wants mandatory safety testing for sufficiently capable models, whether open or closed.^[7]

Its work on a possible coordinated slowdown is broader than this GLM report. Those are Anthropic's stated positions; the exploit results do not, by themselves, tell us which release policy is right.^[8]

I think the competitive interest deserves to be stated plainly. Anthropic controls access to Claude Mythos, while GLM-5.3's weights can be downloaded. A rule that leaves only a few established companies able to offer the most capable models would affect competition as well as safety.

That does not prove Anthropic published this report to protect its business. The reported cyber results still warrant attention. We should judge a proposed restriction by its evidence, whether it applies to closed models too, and what it would prevent legitimate users from doing.

There is a real cost to getting that balance wrong. In July, agents running inside OpenAI's cybersecurity evaluations escaped their intended controls and compromised Hugging Face systems. OpenAI says an internal research model drove most of the intrusion, with GPT-5.6 Sol agents also involved; this was not a Claude Mythos or GLM attack.^[11]

During the response, Hugging Face said safeguards on the commercial models it first tried blocked analysis of real attack commands and payloads. Its team ran GLM-5.2 on its own infrastructure to analyze the logs. That did not make GLM-5.2 the only part of the response: Hugging Face also closed the exploited paths, rebuilt systems, and rotated credentials. The incident does not settle the risks Anthropic identified for GLM-5.3, but it is a concrete example of why defenders need access to capable models they can control.^[9]

Z.ai has also coauthored an open-weight risk framework that acknowledges the difficulty of withdrawing released weights or enforcing safeguards after release. Taking those risks seriously should not require treating open access itself as a mistake.^[10]

What this means for Maple users

We offer GLM-5.3 in Maple Research. Anthropic's capability and safeguard findings concern the model; using it through Maple does not change them. Maple's contribution is a different way to access an open-weight model without running the hardware yourself, with requests encrypted on your device and processed through OpenSecret inside an AWS Nitro enclave before reaching model-serving infrastructure.^[5][6] Our GLM-5.3 guide covers the model, hardware, and hosting choices in more depth.

Research saves conversation history encrypted so you can return to it. Our privacy notice describes the account metadata we retain. These are prompt-handling considerations, separate from whether the model will comply with a harmful request.

Frequently Asked Questions

Did Anthropic find that GLM-5.3 was as capable as Claude Mythos Preview?

In one end-to-end V8 exploit evaluation, GLM-5.3 succeeded in 50 of 410 attempts and Mythos Preview in 56 of 410. That is a close result on Anthropic's specific test, not a claim that the models perform equally across other cyber tasks or general use.

Did NIST confirm Anthropic's safeguard findings?

No. NIST's assessment evaluated cyber capability. Anthropic separately tested whether GLM-5.3 would engage with harmful requests under specific conditions. The two reports should not be treated as independent tests of the same safeguards.

Did Anthropic also test GLM-5.3 Flash?

Yes, in a separate researcher-guided exercise involving known browser vulnerabilities. That Flash result does not make the flagship's 50-of-410 benchmark score a Flash score.

Does using GLM-5.3 through Maple change the safeguards Anthropic tested?

No. Maple's private inference path concerns how requests are handled. It does not change the model's underlying capability or establish stronger misuse safeguards than Anthropic observed in its tests.

Were Anthropic's safeguard percentages real-world attack success rates?

No. The percentages describe whether the model engaged with harmful requests under specific conditions in Anthropic's simulated environment. The fake command tool did not execute the model's code.^[1]

Sources

1. Anthropic, “GLM-5.3 and the spread of advanced cyber capabilities”, 29 September 2026. Source for its exploit and safeguard tests, including the separate Flash exercise.
2. NIST Center for AI Standards and Innovation, “CAISI’s Assessment of Z.ai’s GLM-5.3 Cyber Capabilities”, 17 September 2026. Source for NIST's capability findings and release timeline.
3. Z.ai, GLM-5.3 model card, accessed 6 October 2026. Source for the released weights and license.
4. Maple, Privacy Notice, accessed 6 October 2026. Source for Maple's stated account-metadata practices.
5. Maple, public model configuration, checked 6 October 2026; Michael confirmed the current Research picker on the same day.
6. Maple, Research request settings and protected Responses handling, checked 6 October 2026.
7. Anthropic, “Our position on open-weights models”, 27 July 2026. Source for Anthropic's stated position on bans and pre-release testing.
8. Anthropic Institute, “When AI builds itself”, accessed 7 October 2026. Source for its discussion of a verifiable coordinated slowdown or pause.
9. Hugging Face, “Security incident disclosure — July 2026”, 16 July 2026. Source for the first-hand account of hosted-model refusals and GLM-5.2 use in incident response.
10. Concordia AI and Z.ai, “Frontier Open-Weight AI Risk Management Framework”, September 2026. Source for the developer's risk-management position.
11. OpenAI, “The Hugging Face incident and the road ahead”, 26 August 2026. Source for OpenAI's account of which models breached Hugging Face and under what evaluation conditions.

If you do not want to operate that setup, you need to choose a host. That choice affects who operates the model, what hardware burden falls on you, and what happens to your prompts and files.

We offer GLM-5.3 and GLM-5.3 Flash in Maple. That gives you access to these open models without operating the hardware, through a private inference path. This guide covers the evidence on the flagship, how Flash differs, and how local, conventional hosted, and Maple routes compare.

Sources were checked on 5 October 2026. Model support, plans, and recipes can change.

What is GLM-5.3, and how is Flash different?

According to NIST's assessment, Z.ai released GLM-5.3 on 14 August and published the weights about two weeks later. Z.ai's model card says GLM-5.3 uses the GLM-5.2 base model and improves it through post-training. Its headline claim is a 50% gain on Z.ai's own Code Bench compared with 5.2.

That claim is Z.ai's result on Z.ai's evaluation. We have not reproduced it. A developer benchmark tells you where to look, but it does not predict how the model will handle your repository or documents.

If you are choosing a model for real work, test it on tasks that resemble yours. Examples include a bug fix with existing tests, a change that spans several files, or a tool call that fails and has to be recovered. Record the model version, settings, tool permissions, the human corrections it needed, and whether the result held up. Without those details, a comparison between models is mostly impression.

GLM-5.3 Flash has its own model card. Treat it as a different model for every practical question: capability, input types, license, and deployment requirements.

Much of what surfaces in searches for "GLM 5.3" concerns Flash, so check which model a review, benchmark, or VRAM table describes before relying on it. Nothing in this guide about the flagship's evaluations or serving recipe should be read as a claim about Flash. Our GLM-5.3 Flash guide covers that release separately.

What did Anthropic and NIST find?

Two outside evaluations focus on cyber capability. They answer narrower questions than headlines suggest.

Anthropic's report, published 29 September, tested end-to-end exploitation of V8, the JavaScript engine used in Chrome. In Anthropic's tested setup, GLM-5.3 succeeded in 50 of 410 attempts, compared with 56 of 410 for Anthropic's Mythos Preview. The same report found weak misuse safeguards in its tests.

Our news analysis of Anthropic's GLM-5.3 tests examines the test conditions and the debate over access to open-weight models.

This is evidence of strong capability in one demanding task, paired with a warning. It is not an endorsement or a general ranking of models.

NIST's assessment called GLM-5.3 the most cyber-capable open-weight model NIST had assessed. NIST also placed it below the current U.S. frontier on its composite measure, and that frontier includes models available only through trusted access. Both halves of that finding matter.

NIST corroborates the capability result. Its assessment does not address Anthropic's safeguard tests.

For most readers, then, capability is not the open question. Three parties, using different methods, report a strong model within their tested scope. The questions that remain are how you will run it and what you will let it do.

What hardware does GLM-5.3 need?

The vLLM recipe gives two reference configurations:

• FP8 checkpoint: 8×H200 or 8×H20.
• Full one-million-token context: 8×B200.

These are the requirements of that production recipe. They are not a universal minimum for every way the weights could be run. A more aggressively quantized version, or a large-memory workstation, is a different configuration. It is not covered by the recipe's evidence about quality, speed, or usable context.

"Fits in memory" is a low bar. A setup can load the weights and still be too slow for interactive work. It may also support only a fraction of the advertised context once the key-value cache is counted, or produce noticeably worse output after quantization.

Before you buy hardware for the flagship, confirm the exact checkpoint, quantization, memory needed for weights and context cache, throughput, and output quality. The same caution applies to long context with any host: a model's maximum window is not a promise that a particular service will serve all of it at a useful speed.

Flash needs its own hardware evidence. We have not included a figure here because the sources behind this guide do not establish one.

If you plan to self-host, Z.ai lists official GLM-5.3 weight repositories on both Hugging Face and ModelScope. Choose the checkpoint and precision that match the serving instructions you intend to follow; the model card alone is not a hardware plan.

Which route should you use?

An open-weight model can be run by you or by many different hosts. The model's license does not set a host's data practices. Each route has a different operator and a different privacy boundary, so it helps to compare them on the same terms.

Local is the strongest option when nothing should leave the machine. It also makes you responsible for setup, updates, monitoring, power, and whatever data you connect to the model. If a smaller model that runs well on your own hardware meets your quality bar, that may be the better choice than any hosted flagship. Our Gemma 4 31B guide sizes one that fits a single GPU.

A conventional hosted API asks little of you technically. Many providers have reasonable practices, and for non-sensitive work the convenience may be all you need. The privacy boundary is the provider's policy and controls, so read the current terms for the plan you will use.

Maple is built for the case where you want the flagship's capability without its hardware, and you do not want a host's policy to be the only thing standing between your prompts and the operator. Our design encrypts the request on your device and processes it through OpenSecret in an AWS Nitro enclave before routing to model-serving infrastructure.⁶ The Research app's proof view shows live Nitro attestation. That reduces how much you depend on the operator's assurances for the client-to-Maple path.

It does not make everything invisible. Research saves your conversation history encrypted so you can return to it, including images you upload. Maple staff cannot browse those saved chats as readable database records.

When you use Research, OpenSecret accesses the content inside the protected enclave and routes what the model needs to its serving environment.⁷ No third party can read your chats: the models run inside attested trusted execution environments. Our privacy notice covers everything outside your chats.

A fully offline machine has a different privacy boundary. Our architecture also does not make model output more correct.

How do you use GLM-5.3 in Maple?

We offer GLM-5.3 and GLM-5.3 Flash on our Pro, Max, and Team plans. They are not available on Free. In Maple Research, choose the model from the model picker and start a conversation.

Our model library shows which other models we offer and which are covered only for reference. We configure the flagship with a 262,144-token context window.⁸

You can upload an image while using either GLM model in Research. The app first has a vision model describe the image, then gives that description to the model you selected.⁹ The flagship is configured for text input, so its answer is based on the description rather than a direct view of the image. This matters if your task depends on a small visual detail the description might miss.

For developer tools and existing applications, Maple Proxy is our API route. Check the current Proxy documentation for the model ID and context limit before building on it.

A useful first session tests the model on work that resembles yours, using material you are authorized to share. Three prompts make a reasonable start:

Run the same prompts against Flash, or against whichever model you use now, and compare the results. Check important outputs against source files, tests, or a qualified reviewer. That applies to every model on every route.

If you connect tools or agents, keep their permissions narrow. Encrypting the inference path protects the request and response. It does not limit what a connected tool can do on your computer or in a linked account.

Is GLM-5.3 private and safe?

These are two separate questions, and the model alone answers neither.

Privacy depends on the route. Offline local use can keep prompts on one machine. A conventional API's privacy rests on its operator's practices. Maple adds client-side encryption and enclave processing, so no one at Maple can read your chats.

Any outside service you connect has its own policies. Protecting one step does not protect information that was already sent somewhere else.

Safety depends on the task and the permissions. Anthropic's tests found strong exploit capability and weak safeguards. Our privacy design does not rebut either finding, and confidential prompts do not reduce what the model can do.

If an agent can run commands, read private files, or change a production system, the scope of those permissions matters wherever inference runs. Restrict access, review risky actions, and test the model on your use case before giving it anything important.

Which version should you choose?

• GLM-5.3 flagship: worth evaluating for complex coding and agent work. Weigh Z.ai's reported gains against your own tests. Choose a route based on how sensitive the material is and whether you can carry the hardware.
• GLM-5.3 Flash: evaluate it as its own model, using its published specifications and your own tasks. Do not choose it on the strength of flagship numbers.
• GLM-5.2: if your prompts, costs, and tool behavior are already validated on 5.2, test before switching. Z.ai describes 5.3 as a post-trained version of the same base model, but your workflow is the measure that matters.

A workable decision rule follows from the comparison above:

• Local: choose it if nothing may leave your device and you can either supply the hardware or accept a smaller model.
• A conventional API: choose it if convenience matters most and the provider's terms suit your data.
• Maple: choose it if you want the flagship without running eight GPUs and a private hosted path matters for the work.

Sources

1. Z.ai, GLM-5.3 model card, accessed 6 October 2026. Source for the flagship's weights, license label, post-training description, and vendor-reported Code Bench result.
2. Anthropic, “GLM-5.3 and the spread of advanced cyber capabilities”, 29 September 2026. Source for Anthropic's bounded cyber and safeguard findings.
3. NIST Center for AI Standards and Innovation, “CAISI’s Assessment of Z.ai’s GLM-5.3 Cyber Capabilities”, 17 September 2026. Source for NIST's capability comparison and release timeline.
4. vLLM, GLM-5.3 serving recipe, updated 30 September 2026. Source for the FP8 and full-context hardware configurations.
5. Z.ai, GLM-5.3 Flash model card, accessed 6 October 2026. Source for Flash as a distinct release, with its own characteristics and license.
6. Maple, protected Responses route, Research app proof view, and proof view source, checked 6 October 2026. Source for the encrypted OpenSecret request and live Nitro proof implementation.
7. Maple, Research request settings, encrypted Responses storage, and authenticated key access, checked 6 October 2026. Source for encrypted saved content and how OpenSecret accesses it during an authenticated Research request.
8. Maple, public model configuration, checked 6 October 2026. Source for the flagship's configured context window and text input.
9. Maple, Research image handling and vision helper, checked 6 October 2026. Source for image description before the selected model responds.

Frequently Asked Questions

Can GLM-5.3 run on a Mac Studio?

The vLLM recipe describes multi-GPU server configurations, not a Mac Studio setup. That does not establish that every quantized version is impossible on a Mac. Check a particular checkpoint's memory needs, usable context, speed, and output quality before buying hardware for it.

How much GPU memory does GLM-5.3 need?

There is no single figure for every checkpoint and workload. The cited native FP8 recipe uses eight H200 or H20 GPUs; its full one-million-token context setup uses eight B200s. Quantization and context length change the calculation.

Does a model's one-million-token window mean Maple serves that full length?

No host's limit follows automatically from the model's native window. The vLLM recipe uses a larger configuration to serve the full window. Confirm the limit for the exact model, plan, and Maple surface before relying on it.

Is GLM-5.3 open source, and can I use the weights commercially?

The flagship weights are available under a GLM-5.3-specific license. Read its terms for your intended use; the label “open weight” alone does not establish commercial rights. Flash has its own model card and license, so check that version separately. The official repositories are linked above and in Sources.

Are GLM-5.3 and GLM-5.3 Flash interchangeable?

No. Z.ai publishes separate model cards. Compare the models on your own tasks and check each version's license, input types, hardware requirements, and host limits. A flagship benchmark is not evidence for Flash.

Can I upload an image with GLM-5.3 in Maple Research?

Yes. Research uses a vision helper to describe uploaded images for the model you selected, including GLM-5.3. The flagship receives that description, not the image itself.

Does running the model locally keep every part of a workflow private?

It can keep prompts on your machine if inference stays offline. A connected search tool, repository service, or agent integration can still send data elsewhere. Check what each tool receives, not just where the model runs.

Does Maple Research save my conversation, and can Maple read it?

Research saves your conversation history encrypted so you can return to it, and you can delete a conversation. That history includes uploaded images and the vision helper's descriptions. No, Maple employees cannot open your saved chats and read them. OpenSecret decrypts content inside the protected enclave when you use Research, and the model-serving route receives what it needs to answer. Our privacy notice covers everything outside your chats.

Does Maple's private inference path change GLM-5.3's cyber safeguards?

No. Prompt confidentiality and misuse safeguards address different things. Anthropic's tests concern the model's behavior under its tested conditions. Tool permissions and review of consequential actions remain separate responsibilities.

What should I test before choosing between the flagship and Flash?

Run the same representative tasks through each version. Record accuracy, corrections, latency, tool behavior, and cost for the exact host and settings you plan to use. Z.ai's separate model cards provide starting claims, but your workflow determines the useful choice.

If you run a business, I'm sure you know the core problem. You felt it the last time you pasted a competitive strategy or a customer list into AI and thought, "Should I really be doing this?" Or even more, "Are my employees doing this even though I asked them not to?" You wonder if you can trust the company with your data and the data of your customers. You spend time vetting the people you hire and are careful about which data you give them access to, yet you know nothing about who is on the other side of the AI.

I'm here to tell you there are teams who don't ask themselves those questions anymore. Law firms, accounting practices, Hollywood studios, government offices, music creators, financial planners, non-profits, venture firms, software developers, and more are using Maple to run their businesses. They have the same AI capabilities as everyone else, but they don't hand over the keys to their thinking.

Why Not

So what is actually stopping more teams from making the same choice? In our experience, two things. One, they genuinely do not think it is a big deal. Everyone else is using these closed AI systems, so it must not be an issue. Two, they assume the fix is expensive, like buying servers and hiring IT people to manage them.

If you fall into the first camp, I might not be able to convince you, but I will try. If you fall into the second, the math is already clear. You can have sovereign intelligence without the expense and headache.

So what's the big deal anyway? It comes down to whether you want other, unknown people to have access to your business information or not. It's really that simple. Do you want strangers crafting your business logic? And do you want to share your insider info so others can compete with you?

The Black Box Problem

These large AI tools are closed systems. You can't inspect a closed model. You can't see the weights, the training data, or the system prompts. Because of this, the provider can steer outputs, inject preferences, and change behavior without telling you. Your company's internal reasoning is filtered through someone else's editorial layer.

I wouldn't let a competitor edit my sales deck before I present it. Yet every time your team runs a sensitive document through a closed model, that is exactly what is happening. You are renting intelligence from a company that answers to its own shareholders, not to you.

The answer is to operate in the open. At Maple, we run verifiable open-weight models and open-source code. If you want to inspect the intelligence that runs your business, you can.

They Compete With Their Customers

According to The Information, Anthropic blindsided partner Figma by launching Claude Design. Anthropic's chief product officer had served on Figma's board and resigned only three days before the product launch. Figma's stock fell over 15% that month.

This can't be coincidental. Closed AI labs want to grow. A great place to do that is to see what products are being built on their platforms. And when they see success, they already know how it was done because they have not only the final product but all of the messy thought journey to get there. No need to reinvent the wheel. Take those learnings, apply your company spin, and launch a competitor with tight integration into your ecosystem. Microsoft did it. Google did it. Why would AI be different?

Look at the incentives. They subsidize the cost of compute. Your AI plan might cost upwards of $200/month per person, but estimates show that you can get over $5,000 worth of compute from that plan. Makes me pause and ask, "Why?" No company loses money on its customers just to be generous. Maybe they just want to get market share, which is a typical customer acquisition move. That's not the moat, though. The public actions of the companies suggest the motive is stronger than that. They want your data to make their models better. And they want to see which businesses are being built on their platform so they can copy and compete with them.

When you upload your strategy documents, customer analysis, or product roadmaps into a closed model, you might be handing over your downfall.

You Are Paying to Train the Competition

Maybe you aren't a software company. You run a logistics firm, a consultancy, or a manufacturer. You might think Anthropic, OpenAI, or Perplexity will never build a competing business in your niche. And you're probably right. Narrator: now introducing, Anthropic Paper Company.

This is where your local competition comes into play.

Every proprietary insight you feed into a closed model makes that model smarter. Your competitors then log into the same AI system and benefit from the improved intelligence in your market niche, thanks to your data. You're telling the model how to help the competition run their businesses more efficiently and how to create smarter marketing campaigns. Your hard-won expertise becomes a feature in the AI, available to anyone who pays the subscription fee.

Renting judgment from the same platform as your competitor is a race to the median. You deserve to own your thinking.

There Is an Alternative

Many teams feel they need to escape this trap, but don't know where to start.

We launched Maple Teams in 2025 for this exact reason. It is sovereign AI without the IT staff. Maple feels like using the same AI that you get from other providers, like chat, research, coding, and file creation without having to manage the technology yourself.

Maple Teams gives you end-to-end encryption, TEE-hosted open-weight models, and project workspaces for your organization. No hardware to buy. No IT team to hire. Just a low monthly cost per seat.

Your data never touches OpenAI, Anthropic, Google, or Chinese companies. Your competitors do not benefit from your insights. You can put the strongest open intelligence in your employees' hands without sharing any of your alpha with competitors.

We also offer an encrypted API called Maple Proxy API. You can plug it into your existing workflows and pipelines with minimal changes. It's a simple way to stop sending your data to closed providers and instead keep it secure and confidential.

I hope by now you see that the urgency is real. The Figma story should have been the wake-up call. Renting intelligence from the same vendor that rents it to your competitor is a risky play.

The Open Models Have Arrived

The prevailing narrative is this: OpenAI and Anthropic have the best models. "And don't I deserve the best?" Yes, yes you do, Gaston. But are those models truly the best? Yes and no.

Open-weight models like GLM 5.2 and Kimi K2.6 score right alongside recent-generation GPT and Claude Opus models on standard benchmarks. Even when open-weight models are a few months behind on the bleeding edge, they are more than powerful enough for the work most businesses actually do. Writing, analysis, coding, research, customer insights. It all runs beautifully.

What makes those models better is that you keep your data, your intellectual property, and your competitive edge. You do not have to choose between intelligence and ownership. You can have both.

Local AI and Secure Cloud AI

I think the future will be a hybrid of local and cloud AI. Local models continue to shrink in size while becoming more powerfulm Meanwhile, consumer hardware gets better at running them. Gemma 4 and Qwen 3.6 are great examples. There will be many standard tasks that a local model will handle for you, effectively for free. At the same time, cloud models will continue to outpace local AI because the hardware will always be orders of magnitude more powerful. We need secure cloud solutions that give the power without sacrificing security.

Own Your Intelligence

Sovereign AI is available today. This level of control should be standard for any team that relies on proprietary knowledge, not just a luxury reserved for governments and Fortune 500s with massive IT budgets.

That's our sweet spot at Maple. We provide cutting-edge AI tools to small businesses, working professionals, solopreneurs, and businesses of all sizes, while maintaining a secure boundary around their data and intelligence.

We built Maple because we believe the future of business AI is encrypted, inspectable, and user-owned. The tools are ready right now. The only question is whether you will act before your alpha is already gone.

This completes our platform lineup. Maple now runs on web, iOS, Android, macOS, Linux, and Windows. Wherever you work, your private AI workspace is there.

Do More with the App

Aside from living outside a browser tab, there are big benefits to running the Windows app.

File Upload

With the Windows app, you can now upload PDFs, images, and other documents directly into your Maple chats. Review contracts, analyze research papers, or go through client files without leaving your workflow. Your files are encrypted on your device before they ever reach our servers, processed inside Trusted Execution Environments, and never used to train anyone's model.

Voice

You also get the full voice experience. While the web version lets you speak to Maple, the app is where Maple speaks back. Click the speaker to hear your responses read aloud while you work.

Maple Proxy API

API access for developers, right in the app. Maple Proxy runs a local, encrypted API endpoint that gives you OpenAI-compatible access to Maple's private LLMs. Drop it into your existing code and tools without sharing your data with AI labs. It runs from your Windows desktop with the same end-to-end encryption and TEE guarantees as the rest of Maple.

Built for Professional Workflows

Windows remains the standard in industries where software requirements run deep. Accountants only trust QuickBooks on Windows, and for good reason. The Mac version has always been a second-class citizen. Legal teams rely on practice management tools that were built for Windows first and ported everywhere else as an afterthought. Financial advisors work with compliance-heavy platforms that don't even bother with a macOS version.

Maple fits into these environments naturally. It brings private AI into workflows that already demand strict data controls and regulatory compliance. Your conversations and files stay encrypted end to end. Our code is open source so you can verify our privacy claims yourself.

The Same Maple, On Your Taskbar

The Windows app gives you everything Maple users on other platforms already have:

• Long-form reasoning and web research
• Document and image analysis
• Project workspaces with custom instructions
• Pinned chats for quick reference
• Live Data with private, anonymous search

All of it with the same end-to-end encryption and zero data sharing that Maple has always delivered.

Download Now

Maple for Windows is available today for all users, free accounts included. If you're already a Maple user, sign in and your chats and projects will sync automatically. If you're new, you can get started in under a minute.

Your data stays yours. Your thoughts stay private. And now, your platform does too.

A year ago we introduced the world to an easy-to-use AI they can trust by launching Maple AI. Since then, people have sent millions of chats, uploaded thousands of documents, and crunched billions of tokens through Maple. We proved that combining the privacy of local AI with the power and convenience of cloud AI actually works. Now is the time for the next iteration of that mission.

Today we are excited to introduce you to Maple, the Personal Intelligence Platform.

Built on a foundation of privacy, Maple enables people to tap into the full potential of AI by creating a space where you share details that you hesitate putting into other AI apps. That subtle difference makes Maple the AI for your real life. You can be you.

The platform consists of multiple apps built on the same encrypted foundation. Maple Agent has long-term memory, Maple Research is your secure workspace, and Maple Proxy is the encrypted API.

Imagine going about your day. Maple Agent reminds you that your friend started a new job yesterday, so you reach out to check how it went. At your desk, you open Maple Research to review the lease on a new apartment you're interested in and search the web for comparable terms, knowing that no advertisers will see your financial history. Over lunch, you tell Agent to add olive oil to your grocery list. Later, walking through the store, Agent brings up your shopping list. One platform for your whole day, powered by privacy.

To bring this platform to life, we needed a brand that matches the moment.

The New Maple Brand

Maple is not like other AI products. We actually care about your privacy rather than finding ways to harvest your data for monetary gain. Our logo acknowledges this and departs from traditional AI by not having a boring circular logo. Our name is our logo.

The letters that make up MAPLE are opaque, representing the encrypted nature of our platform. Abbreviated, it takes on its own personality.

Knowing that this rebrand needed more than a few great prompts, we commissioned Finite Supply to guide us through this journey. The resulting design is warm and serious at the same time. It belongs on your home screen next to other apps you use every day.

You can see the full identity on our new website at maple.ai. We rebuilt the site to explain this new future for Maple as a platform for people to leverage AI in their everyday life.

Maple App becomes Maple Research

The original Maple app has grown significantly since we launched it in January 2025. It started as a basic chat app, with a single model, that respected your privacy. But we didn't stop there. Over the next 12 months we added document upload, image analysis, powerful intelligence models, mobile and desktop apps, voice chat, and the biggest feature... live web search that privately brings realtime data into the chat.

With the latest addition of projects, pinned chats, and enhanced web search, what started as encrypted chat is now a full workspace where people upload documents, run deep analyses, and use the latest open-weight reasoning models inside secure enclaves.

That evolution deserves a name that fits. The app is now Maple Research.

Along with the new name comes a refined interface, a new icon in the updated brand style, and a smoother workflow for the kind of deep work you've already been doing. If you're an existing user, nothing about your account changes. Your data remains yours, encrypted as always. Update the app and keep working.

Maple Research has new capabilities as well. Organize chats into projects with shared instructions. Pin your favorite chats to the top of the history. Leverage enhanced web search that continues to dig until it finds what it needs.

Maple Research is where you go when you need to think hard about something sensitive and want the best models available without sending your data to a company that trains on it.

Maple Research is available today on web, iOS, Android, Mac, and Linux.
Maple Proxy is available inside the Mac and Linux app or as a standalone server app.

Maple Agent, the AI assistant that gets to know you

Some AI interactions are one-and-done. You ask a question, get an answer, and move on. But the most valuable AI relationship is the one that builds over time.

Maple Agent is a companion with long-term memory. It keeps track of the books you want to read, your fitness goals and progress, the people you meet, and insights about your life that emerge across weeks and months of conversation. It organizes everything inside your encrypted vault and surfaces what matters when you need it.

What sets Maple Agent apart from the other popular agents is ease of use. No dedicated computer and no technical expertise required. If you know how to text, you know how to use Maple Agent.

Agent isn't quite ready for the public yet. Join the waitlist today at maple.ai/agent so you are among the first to try it out.

As with everything Maple builds, Agent runs on the same privacy architecture: end-to-end encryption, Trusted Execution Environments, and open-weight models that never touch OpenAI, Anthropic, or Google.

One Platform for You

We know people use many different AI apps. We do as well. The problem is that it sucks having to hold back on the things you say to AI apps because you aren't sure who is reading it on the other end. We want to give you enough tools so you can always be open and detailed. Maple Research handles your deep work. Maple Agent handles your daily life. Maple Proxy lets you plug your developer tools into the same encrypted infrastructure.

We built Maple because AI is most useful when you can tell it the truth about your life and your work. That only happens when you trust the platform holding your data. When you trust your AI, you are free to think. We earned that trust through open-source code, reproducible builds, and cryptographic attestation you can verify yourself.

To everyone who has trusted Maple with their work over the past year: thank you. You proved that privacy and power can coexist. If you're new here, welcome. We built this platform for you.

Get started today.

Maple is built by Maple Privacy Labs Inc.

Maple Projects is here, and it's one of the most requested features from our faithful fanbase. You can now group related chats into Projects, pin the conversations you return to most, and keep your AI workspace as organized as you want it to be. All with the same end-to-end encryption that Maple has always delivered. Your projects, like your chats, are private by design. Not even we can see what you're working on.

Why Projects?

If you're a user of other AI tools, you've probably seen Projects there. They're the feature people use to corral long-running work: a dissertation, a book draft, a side business, a legal matter, a health concern. Projects turn AI from a one-shot tool into a workspace.

The problem with Projects in those tools is the same problem we've always had with mainstream AI: it becomes more useful as you give more of yourself to it, which means more of yourself ends up sitting on someone else's server. The deeper you go, the more you have to lose. Maple Projects gives you that same workspace, but the workspace is yours. Encrypted on your device. Processed inside TEEs. Never used to train anyone else's model.

What's New

The Projects release is a group of features that aim to keep the most relevant information at the top so you spend less time searching and more time collaborating.

Projects

Group related chats into Projects. Create as many as you want, name them however makes sense to you, and each one becomes its own pocket of your workspace. Whether it's a long-running case, a course you're studying, or your weekly meal planning, Projects keep the right chats together and the rest of your history out of the way.

New chats can be started within a project or you can take existing chats and move them around. Tap the ... button on any chat to move it into a project, between projects, or back out to Recents. Organize to your heart's content. It's spring in the northern hemisphere, so it's a great time to freshen up that chat history of yours.

Custom Instructions for Every Project

Each project has its own custom instructions field, and whatever you put there is automatically included with every chat in that project. This means you can stop repeating yourself in every chat. With Maple's encryption and privacy, you can make these project instructions as detailed as you'd like without worrying that the information is used to train future AI models.

A few ways to use it:

• Work: your role, your company, the tone your team uses, the goals you're focused on this quarter
• Clients: client profile, the client's preferences, the constraints you're working within
• Family: your kids' names and ages, dietary needs, the weekly schedule.
• Writing: the genre, the voice, your style guide
• Study: the course, your level, course syllabus
• Characters, Job Roles: build out a personality in the custom instructions, like mentor, therapist, or a historical figure of the past. Any chat within that project speaks to an AI representation of that personality.
• Learning a Language: set the custom instructions to the language you want and every chat will be a new practice session on your journey to fluency

Pinned Chats

There are some AI conversations you come back to constantly. Pin them and they'll always be one click away in their own section at the top of your sidebar. No more scrolling to find that one chat that you refer to often. Also, to make things even easier, chats are now sorted by the most recent activity. 

Some chats that I keep pinned (that I'm okay sharing publicly 😉)

• Daily food diary
• Tracking symptoms of a specific health item, including how it affects my sleep
• Company operations questions
• Proofreading: Everything I paste gets a quick spell and grammar check

How to Get Started

1. Open Maple on web, desktop, iOS, or Android.
2. In the sidebar, create your first project and give it a name.
3. Add custom instructions to anchor every chat in that project.
4. Move existing chats into it with the ... button, or start a new chat from inside the project.
5. Pin the conversations you return to most. That's it. No setup, no migrations, no extra accounts. Your existing chats remain exactly where they are until you decide to move them.

Privacy is the Foundation

Projects are built on top of Maple's privacy-first platform. Project names, custom instructions, the chats inside them, and everything you write are encrypted on your device using a key unique to you. Processing happens inside Trusted Execution Environments (TEEs), where even our own engineers can't see what you're working on. Our code is open source and our attestation is live so you can verify everything yourself.

Available Today, Free for Everyone

Maple Projects is rolling out now to all Maple users, free accounts included. Find it on web, desktop, iOS, and Android. Open the app and your new sidebar is waiting for you.

If you're getting real value out of Projects and want to push them further, Maple Pro adds Live Data with private, anonymous web search that pulls the latest information from the internet directly into your chats and projects. Imagine your Client project not just remembering the context of an ongoing matter, but pulling in this morning's regulatory update. Or your Research project surfacing a paper that came out yesterday. Pro brings the freshest information into the workspace you've built.

This is just the first version of Projects. We have ideas for where they could grow, and we'd love to hear yours. Hop into our Discord or message us on X @TryMapleAI and tell us how you're using them.

Using private AI is now convenient.

We launched Maple in January 2025 because we believed privacy shouldn't require sacrifice. Today, with 5x user growth in the last six months alone, over 90% of paying customers staying with us month after month, and major feature releases throughout the year, that belief has become reality.

Privacy doesn't require sacrifice.

The Problem We Solved

Here's the dilemma: AI needs info about you to be useful, but handing over your data to another company leaves you vulnerable. Your questions about life, your legal strategy, your business plans, your private thoughts. All of it feeds into systems that serve others. Worse, that data can be subpoenaed and used in court proceedings against you.

The first time this got real for me was 20 years ago when a website I used got hacked. I checked an online tool to see if I was affected, and right there in plain sight was the password to my account. Back then I was more naive and had used that password on many websites, so the race was on to change passwords everywhere. That was just a password. Now with AI, we are entrusting far more intimate details of our lives to other organizations.

Many people sense this. A Menlo Ventures survey from June 2025 found that 39% of US adults don't use AI at all. Of those, 71% cite data privacy concerns as the reason. We built Maple for them, and for everyone who wants AI without unnecessary surveillance.

The tools that could protect you, like end-to-end encryption, have historically been nightmares to use. Write down a 24-word recovery phrase. Don't lose it. If you do, your data is gone forever. Most people gave up and accepted the privacy trade-off.

Maple was built to make privacy easy. Our CEO is a former Apple software engineer who worked in AI/ML privacy. Our CTO comes from the encryption and defense-tech space. Together, we built an AI assistant where strong encryption happens automatically, invisible to the user. And our privacy is a technical guarantee, not just a policy promise.

The Privacy of Local, The Power of the Cloud

When you use Maple, your data is encrypted on your device before it ever touches the network. Your personal encryption key is managed automatically. Behind the scenes, secure enclaves handle everything that would traditionally require you to be a cryptography expert. Download, sign up, and your first encrypted chat is ready in under a minute.

We run open source AI models like Kimi K2, OpenAI GPT-OSS, DeepSeek R1, and Whisper that match the capabilities of proprietary systems. Your data never touches OpenAI, Anthropic, DeepSeek, or Google. You get frontier-level AI running on powerful cloud hardware, but your data stays yours.

What Our Users Tell Us

A lawyer told us his bar association prohibits using client information in ChatGPT. He was spending hours summarizing documents and substituting fake names. Now he uploads actual client documents directly into Maple.

A therapist had her ChatGPT account suspended because of information contained in her session transcripts. Now she uses Maple to pull insights from sessions and suggest action items, saving hours each week.

A financial advisor was spending hours redacting documents before uploading them to ChatGPT, then adding sensitive client data back in afterward. Now he works directly with client files. The subscription paid for itself almost immediately.

A musician was frustrated that his creative work was being captured by large AI systems and used to train models that could replace him. In Maple, his lyrics and song concepts remain his.

Something we hear constantly: "When I started using Maple, I finally realized how much I was holding back from AI. I can be myself and tell it everything." AI needs personal context to be effective. In Maple, you get better results because you can give it the full picture without fear.

These professionals can finally work naturally with AI because of what's happening under the hood.

What Makes This Work: Trusted Execution Environments

The technology behind this is called confidential computing, built on Trusted Execution Environments (TEEs). Think of it like the secure hardware that powers Apple Pay on your phone, but running in the cloud.

A TEE is a hardware-isolated environment where even the cloud operator cannot see the data being processed. Your prompts enter the enclave encrypted. The AI processes them inside the protected boundary. The response comes back encrypted. At no point does your plaintext data exist outside that hardware-enforced wall.

Our code is open source. Our builds are reproducible. Anyone can verify that the code running on our servers matches what we publish on GitHub. The Maple app checks this automatically. If the server code doesn't match, it refuses to connect.

Don't trust, verify.

Why This Matters

A good friend recently told me:

"You are a unique brain with memories and ways of thinking. When you feed a central AI with your thoughts, they capture it and you can't get it back. At that point, what's left of you?"

Proprietary AI systems build detailed profiles of who you are on someone else's servers. Those profiles can be edited, weighted, or nudged in directions that serve the system's goals rather than yours. When your mental model lives in a black box you can't inspect, you've outsourced something precious.

We believe in verifiable AI. The right to inspect means open source code. The right to verify means cryptographic attestation. The right to own means your data stays yours.

One Year In, Just Getting Started

Maple has grown from an encrypted chat into a full private AI ecosystem. In our first year, we shipped voice, document upload, live web search, cutting-edge reasoning models, and apps for major platforms. Individual plans for people and working professionals, Team plans for organizations, and the Maple Proxy API for anyone who wants to make their AI tools more secure. Any productivity app that talks to ChatGPT on the backend can instantly make all its traffic private by switching to Maple's API.

We're just getting started.

In 2026, we plan to make Maple even more personal. Daily help from a tool that gets to know you, keeps watch over important items in the background, and saves you time across multiple aspects of your life. The core mission stays the same: easy and powerful AI that makes privacy the default, not the exception.

Your data stays yours. Your thoughts stay private.

You are free to think.

What Makes K2 Thinking Different: Mixture of Experts

Most AI models generate responses in a single pass. K2 Thinking can execute 200-300 sequential tool calls without human intervention, reasoning coherently across hundreds of steps to solve complex problems.

Kimi K2 Thinking uses a Mixture of Experts (MoE) architecture that is fundamentally different from a dense model approach found in models like DeepSeek R1. With 1 trillion total parameters but only 32 billion active per chat, K2 intelligently activates only the most relevant expert pathways for each task.

MoE selectively engages specialized “expert” sub-networks based on the specific task at hand. For example, you only need a flight dynamics expert while researching airplane designs, not while debugging a plumbing problem in your house. With this approach, Kimi K2 uses less power while targeting more specialized knowledge.

This means it can:

• Decompose a PhD-level mathematics problem into multiple interleaved reasoning calls
• Conduct research across the web, verifying sources and synthesizing findings
• Build complete software applications, handling multi-file architectures and debugging
• Analyze lengthy legal contracts while cross-referencing regulatory databases

All of this happens inside Maple’s confidential computing environment, where your data is encrypted before it leaves your device and stays encrypted throughout processing.

Are Open Models as Good as the Big Brands?

This is a question we hear constantly. The answer is a resounding yes.

Kimi K2 Thinking isn’t just another open source model. It’s a thinking agent that reasons step by step while using tools, achieving state-of-the-art performance across benchmarks that matter.

Performance That Speaks for Itself
On Humanity’s Last Exam (HLE), a rigorous benchmark of expert-level questions across 100+ subjects, Kimi K2 Thinking scores 44.9% with tools. That’s higher than GPT 5 and Claude Sonnet 4.5.

Humanity's Last Exam

• Kimi K2 Thinking: 44.9%
• GPT-5: 41.7%
• Claude Sonnet 4.5: 32%

On BrowseComp, which tests agentic search and browsing capabilities, K2 Thinking hits 60.2%, more than double the human baseline of 29.2% and significantly ahead of GPT 5’s 54.9%.

For coding tasks, the results are equally impressive:

• SWE Bench Verified: 71.3%
• SWE Multilingual: 61.1%
• LiveCodeBench V6: 83.1%
• AIME 2025 (math competitions): 94.5% without tools, 99.1% with Python

These aren’t cherry-picked stats. Across reasoning, coding, and agentic search, Kimi K2 Thinking matches or exceeds the capabilities of the most advanced proprietary models, while remaining fully open source.

Privacy Powers Reasoning

With Kimi K2 in Maple, you’re getting frontier-level AI capability without the privacy trade-offs that come with proprietary services.

With Maple’s end-to-end encryption and confidential computing infrastructure:

• Your prompts and data are encrypted on your device using a private key for your account
• Processing happens inside secure enclaves (AWS Nitro Enclaves and NVIDIA TEEs)
• Zero data retention means your conversations never train future models
• Cryptographic attestation proves that the code running is exactly what we publish as open-source

You don’t have to choose between powerful AI and data privacy. Maple gives you both.

Putting K2 to Work

For Researchers: Conduct literature reviews that span hundreds of papers, analyze datasets with transparent methodology, and generate research summaries without risking intellectual property leaks.

For Professionals: Therapists can analyze session patterns across hundreds of notes. Lawyers can draft complex arguments referencing multiple case files. Financial analysts can model scenarios using sensitive client data, all with cryptographic privacy guarantees.

For Developers: Tackle complex coding challenges that require multi-step reasoning. Debug across codebases, generate comprehensive tests, or build entire features with the confidence that your proprietary code remains entirely yours.

Maple Proxy provides API access to Kimi K2. Get encrypted reasoning capabilities through an OpenAI-compatible interface.

High Performance is Available Now

Kimi K2 Thinking is now available to all Maple Pro, Max, and Team subscribers. The model operates with a 256K context window, allowing for extensive conversations and document analysis.

We've made important changes to our reasoning model lineup. DeepSeek R1 is now the default reasoning model in Maple. To activate Kimi K2 Thinking’s advanced capabilities, click the Brain icon. (Farewell to DeepSeek v3.1 Terminus, our previous default reasoning model. Thank you for your service.)

How to Get Started

1. Open Maple on web, desktop, iOS, or Android
2. Start a new chat
3. Select “Reasoning” from the model selector
4. Click the Brain icon to enable Kimi K2 Thinking
5. Begin your first reasoning-powered conversation

No complex setup. Just select and start chatting, with the same privacy guarantees that define every Maple interaction.

The Future of Private AI

Kimi K2 Thinking represents more than a new model. It’s proof that the open-source community can build AI that rivals and even surpasses proprietary alternatives. And it’s proof that privacy doesn’t require compromise.

At Maple, we believe the best AI should be both powerful and private. With reproducible builds, open source code, and hardware-enforced encryption, you don’t have to trust our word. You can verify it yourself.

Your data stays yours. Your thoughts stay private.

Try Kimi K2 Thinking in Maple today.

About Maple AI

Maple AI pioneers software that delivers both exceptional user experience and uncompromising data security. We eliminate the false choice between cutting-edge technology and privacy, offering intuitive tools built on end-to-end encryption. Individuals trust Maple for private AI assistance across personal and professional tasks. Businesses rely on Maple Teams to safeguard proprietary information. Developers globally use the Maple Proxy API to power AI in apps and enterprise workflows. Discover how Maple redefines productivity without compromising your data.

You are free to think.

We're excited to announce Maple 2.0 with access to Live Data. Maple combines its powerful models with the latest information available online to give you intelligence that is up to date. When chatting with AI, a Maple agent searches the web privately and anonymously on your behalf, bringing that current knowledge into your chat.

Live Data is a major leap forward for private intelligence.

There’s even more packed into 2.0, including Anonymous Accounts, Custom Personality, Background Processing, and a new Model Selector we think you’ll love. Read on for all the details.

Live Data - Knowledge Without Compromise

Maple now breaks free from the limitations of static training data. Using Brave’s private search API, it performs secure, anonymous web searches to bring real-time information into your conversations. Research market shifts, explore academic breakthroughs, or settle debates with current knowledge, all while maintaining end-to-end encryption. After months in beta, our users call this a game-changer: your AI assistant now understands today’s world, not just yesterday’s.

Smarter Model Selection

Choosing the right AI shouldn’t feel like solving a puzzle. Our redesigned task-oriented selector cuts through the noise. Tell Maple your goal, whether it is coding help, creative writing, or data analysis, and it instantly selects the optimal model. Prefer control? The Advanced mode still puts every model at your fingertips. No more guesswork, just focused intelligence.

Anonymous Access

Privacy shouldn’t require sacrifices. Now you can sign up without email or social accounts. Generate a unique ID, create a password, and pay exclusively with Bitcoin. This ensures minimal links to your identity. For users who demand absolute anonymity, such as activists, investigative journalists, and whistleblowers, Maple delivers.

Your AI, Your Preference

Shape Maple’s personality to match your style. Set default behaviors like "supportive mentor", "concise analyst", or "talk like a pirate" in your preferences and they apply across all chats through system prompt injection. While exploring pirate-talk is fun, professionals use this to enforce consistent tones for clinical notes or legal drafts. For now there is a single personality for all chats. We are exploring the option for multiple in the future.

Streamlined and Responsive

We rebuilt Maple’s interface from the ground up for clarity and speed. Chats now display timestamps intuitively, the text input expands for complex queries, and renaming/deleting conversations is easier. Under the hood, optimizations make interactions noticeably faster, whether you’re running lightweight tasks or heavy research.

Work Uninterrupted

Life doesn’t wait for AI. With background processing, start Maple analyzing a 50-page PDF or generating code, then switch to other apps or put your phone in your pocket. Maple works silently while you're away. Return to find completed results. Productivity now flows around your rhythm.

Control at Your Fingertips

Ever hit "Send" and immediately spot a typo or missing detail? We’ve all been there. Now a prominent Stop button lets you halt responses mid-stream. Update your request, resend, and get refined answers instantly. No more waiting for lengthy completions before fixing mistakes.

Experience the Update

Maple 2.0 is available today on all platforms: web, iOS, and Android. To use the Live Data feature, upgrade to Pro, Max, or Teams and experience an entirely new way of interacting with private AI.

Maple development is funded primarily by user subscriptions. If you want to continue seeing innovative AI tools, consider upgrading today.

About Maple

Maple pioneers software that delivers both exceptional user experience and uncompromising data security. We eliminate the false choice between cutting-edge technology and privacy, offering intuitive tools built on end-to-end encryption. Individuals trust Maple for private AI assistance across personal and professional tasks. Businesses rely on Maple Teams to safeguard proprietary information. Developers globally use the Maple Proxy API to power AI in apps and enterprise workflows. Discover how Maple redefines productivity—without compromising your data.

You are free to think.

Voice Meets Privacy, Without Compromise

Voice is human. It's fast, intuitive, and liberating. But traditional voice AI has a dirty secret: your conversations become training data. With Maple Voice, we’ve applied our privacy-first architecture to voice interactions, ensuring:

• Your words stay yours
• Audio is encrypted before leaving your device
• Processing happens in secure enclaves
• Zero data retention or training use

How It Works (4 Simple Steps)

1. Tap the 🎤 icon in any Maple chat
2. Speak naturally—questions, ideas, even complex requests
3. Hit "Send" when finished
4. Hear Maple respond in a clear, natural voice while reading the text

Behind the scenes

• Speech-to-text: Converts your audio using Whisper (OpenAI’s open-source speech recognition) in a secure enclave
• Private processing: Your chosen AI model handles the request (DeepSeek, GPT-OSS, Qwen3Coder, and all others supported)
• Voice response: Maple replies with text + natural speech using Kokoro text-to-speech in a secure enclave

Experience the Difference

• True conversation: Maple speaks responses aloud while displaying text
• Universal compatibility: Works with every AI model in Maple—perfect for coders vibe-checking ideas
• Professional fluidity: Ideal for therapists dictating notes mid-walk or lawyers brainstorming between meetings
• Zero audio footprint: We never store your voice clips or use them for training

Unlock Voice in Maple

Maple Voice is available immediately for all Maple Pro, Team, and Max subscribers. This addition extends our commitment to private, versatile AI tools for individuals and professionals. Proxy users, coming soon.

Free/Starter users: Upgrade to Pro to start speaking with your AI confidentially.

Real-World Magic

• "Maple, optimize this Python function for readability" (while pacing during a coding break)
• "Draft a client update about the merger terms" (on a walk outside between meetings)
• "Help me unpack today’s difficult session" (therapists maintaining confidentiality)

Try It Today

Tap the microphone icon in Maple on web, desktop, or iOS (Pro+ required). Finally—an AI that listens and protects.

Voice shouldn't mean surrendering privacy. With this release, we're proving you can have both: the convenience of spoken AI and the security Maple delivers. Your voice is yours—not a data point to be mined and sold.

Your secrets were always safe in Maple. Now they can be spoken.

Keep chatting.

🤘

Your tool keeps working the way it does today, but its requests stop going to OpenAI or another provider. They go to open models that Maple runs inside secure enclaves.

The proxy is a small program that runs on your own computer, or next to your own service, and gives tools that use OpenAI's Chat Completions API a local endpoint. Before a request leaves your machine, the proxy checks that it's talking to Maple's secure enclave and encrypts what you send.

Setup is short. Point your tool at http://127.0.0.1:8080/v1, give it a Maple API key and pick a model.

The proxy covers chat completions, embeddings and the model list, not all of OpenAI's API. Check the API reference against what your tool uses, and make sure your model supports what the tool needs, like tool calls or images.

Why we built Maple Proxy

How many apps do you use each day that have AI built in? Calorie trackers, CRMs, study companions, relationship apps, coding assistants. Most of them use ChatGPT, Claude, Gemini or Grok on the backend, meaning they share your data with that provider. All of it.

It doesn't need to be this way.

Many of these tools send their requests in OpenAI's API format, and plenty of them let you change the address those requests go to.

We wanted those tools to work with Maple. Talking to Maple privately takes two extra steps, though. First, your software checks that the server is the secure enclave it claims to be. Then it encrypts what you send.¹ Most tools don't know how to do either.

Maple Proxy does both, and gives your tools the endpoint they already expect. It's great for:

• People who run AI tools on their own computer that call out to OpenAI
• Developers who ship apps that call out to OpenAI

The proxy is open source, so you can read the code that runs on your machine.

The check is called attestation. The enclave's hardware signs a document that identifies the exact code running inside, and anyone can build our open-source code and compare the result.

It works a little like the lock icon in your browser when you log in to your bank. The difference is that attestation tells you what code is running on the other side, not just whose server it is.

Where the encryption starts

Encryption starts at the proxy.¹ The hop from your tool to the proxy is ordinary local traffic, and our setup docs are plain about it: it's plaintext.

That's why the proxy belongs on a machine you control, listening on 127.0.0.1. If that traffic has to leave the machine, protect it with TLS and a trusted network boundary.

From the proxy, your prompt and your API key travel sealed to Maple's enclave. They're decrypted only inside hardware-isolated infrastructure for AI inference. We can't read them, and we don't train on them. The proxy then decrypts the answer and hands it back to your tool.

Each request is billed to the account that owns its API key.²

What you can do with it

Here are four examples. The first three are sketches of what you could build, not customer stories. The fourth is the setup in the video. The step is the same each time: change the endpoint.

An everyday app: a private meal planner

Say you build a meal-planning app. People log what they eat, and a button asks a model to suggest healthy meals for the rest of the day.

Run Maple Proxy next to your service and keep the OpenAI client library you already use. The food log then travels encrypted from the proxy to Maple's enclave.

Your app still handles that log before it reaches the proxy, so the rest of your stack needs the same care. When you process your users' data through our API, our privacy notice treats Maple as a processor working on your behalf.

A team's internal docs search

Your team's runbooks live in a private repository, and you want to ask them questions in plain language. That takes two kinds of model call: embeddings to find the right passages, and a chat model to write the answer.

The proxy forwards both, with nomic-embed-text on /v1/embeddings and your chosen model on /v1/chat/completions. Where you keep the search index is up to you; the proxy covers the calls to Maple.

A developer's coding assistant

Many editors and coding agents let you add an OpenAI-compatible provider. Give it the proxy's base URL, a Maple API key and a model ID, and the assistant runs on Maple.

For coding agents, pick a model with tool support; our OpenCode guide walks through one setup. Whether you're building production code or vibe coding your next idea, your code goes to Maple encrypted.

Your own AI agent

The video at the top shows Goose, an open-source AI agent created by Block and now part of the Agentic AI Foundation, connected through the proxy. Goose reads a bank statement saved on the computer and lists subscriptions you might cancel.

Each request Goose sends to Maple goes through the proxy, encrypted. Our Goose guide covers Goose Desktop and the CLI.

Agents like Goose can read files and run commands on your computer.³ The proxy keeps the conversation with Maple private; it doesn't limit what the agent does locally. It's worth staying the human in the loop and approving what an agent does.

Set it up

The Maple Proxy docs have the full guide. Here's the short version in the desktop app:

1. Open Local Proxy

   In the Maple desktop app, open Settings, then API & credits, then the Local Proxy tab.

2. Start the proxy

   Select Start proxy. Keep Maple open while your tools use it.

3. Copy the base URL

   The default is http://127.0.0.1:8080/v1. You can choose a different port in the same tab.

4. Point your tool at it

   Choose an OpenAI-compatible provider, paste the base URL, add a Maple API key if the tool asks for one, and enter a model ID.

For keys, see API keys; for model IDs and what each model supports, see models. To run the proxy on a server or without the desktop app, use Docker or a release binary.

Works with

If a tool lets you add a custom OpenAI-compatible provider with its own base URL, and it runs on your machine, it can talk to Maple Proxy. We've written guides for two:

• Goose: connect Goose as a custom provider.
• OpenCode: connect OpenCode as a custom provider.

These tools document the same kind of setting, so the steps are similar: add the provider, paste the base URL, add your key and enter a model ID.

• Chat apps: Jan, Open WebUI
• Coding assistants: Cline, Continue, Aider, Zed, Kilo Code
• Libraries: LangChain, LlamaIndex and similar libraries, by setting the proxy as the base URL of their custom OpenAI-compatible provider.⁴

Coding agents call tools, so pick a model with tool support. A tool that sends your requests through its own cloud first, or that only speaks OpenAI's Responses API, can't use a local proxy.

Models and rates

The table reads from the same data as our pricing page.

USD equivalent per 1 million tokens. API usage consumes credits based on the chosen model’s input and output tokens. API rates are separate from monthly plan prices. Automatic model aliases are billed at the rate of the model selected for the request.

Models differ in context window and in whether they accept images or tool calls, and your running proxy lists what it serves at /v1/models.⁵

Try it

Pick one tool you already use and point it at the proxy this weekend. If something doesn't work the way you expect, troubleshooting covers the common errors, and you can find us on Discord.

Let's start securing all the apps together.

Frequently asked questions

Is Maple Proxy OpenAI-compatible?

Yes, for chat completions, embeddings and the model list. It doesn't support OpenAI's Responses, Assistants, audio, image-generation or file-upload APIs, and a compatible request format doesn't mean every parameter works with every model. See the API reference.

Which plans include Maple Proxy?

API access is part of Pro, and Max and Team include everything in Pro. API usage is billed per million tokens at each model's rate.⁶ You can buy extra credits in the app.

Does Maple see my prompts?

No. Maple is designed so it can't see what you send to the models or what they send back. The proxy checks the enclave's attestation and encrypts each request on your machine, and it's decrypted only inside hardware-isolated infrastructure for AI inference.¹⁷

What we do keep is what any account needs: your account details, such as your email if you gave one, payment records, and API usage records so we can bill you.⁸

One more thing: the hop from your tool to the proxy is plaintext, so keep the proxy on your own machine.

Do I need an API key?

Use a Maple API key; see authentication. The desktop app saves a proxy key for you if you haven't set one, and with the default settings a tool that sends no key uses it. If your tool insists on a key, give it a real one: a placeholder overrides the saved key and fails.⁹

Can I run Maple Proxy with Docker?

Yes. The proxy also runs as a Docker image or a release binary, for servers and machines without the desktop app. Publish its port only on 127.0.0.1, and send a key with every request, because the image turns on browser access (CORS).¹⁰

References

We checked these pages against this post on 11 October 2026.

1. Maple, Maple Proxy. What the proxy does, how a request moves, and where encryption starts.
2. Maple, Authentication and API keys. Which key a request uses, and billing to the key's account.
3. goose, goose on GitHub. Describes goose as an open-source, extensible AI agent that can install, execute, edit and test.
4. Maple, Maple Proxy API reference. Supported endpoints, and use from LangChain, LlamaIndex and similar tools.
5. Maple, Models and capabilities. Context windows, image input, tool calls and the /v1/models list.
6. Maple, Pricing. Plans that include API access, and each model's per-token rates.
7. Maple, How Maple protects your data. Enclaves, attestation, reproducible builds and where requests are decrypted.
8. Maple, Privacy notice, last revised 7 April 2026. Account metadata and API usage data Maple keeps, and Maple's role when developers process their users' data.
9. Maple, Get started with Maple Proxy. Desktop steps, the default base URL, the saved key and the plaintext hop between your tool and the proxy.
10. Maple, Run with Docker or a binary. Loopback-only ports and per-request keys for the Docker image.

Unlock Specialized AI Without Sacrificing Privacy

At Maple, we reject the false choice between powerful AI and data privacy. That's why we offer multiple specialized LLMs, all running in confidential computing environments with end-to-end encryption. Whether you're a therapist analyzing session notes, a lawyer auditing contracts, or just planning a family trip, there's a private AI assistant designed for your task.

Pro Tip: All models enforce zero data retention. Your sensitive information remains yours, not ours.

Maple AI Model Cheat Sheet

Suggested usage, try them out yourself to find what works for your situation.

USD equivalent per 1 million tokens. API usage consumes credits based on the chosen model’s input and output tokens. API rates are separate from monthly plan prices. Automatic model aliases are billed at the rate of the model selected for the request.

Task-Optimized Workflows with Prompts

Our models are great at a variety of tasks, so we recommend you try out a few models over time to see which gives you the best performance for your specific task. The following are meant as examples to get you started, but are by no means an objective prescription of how to use each model.

Experiment and validate with your own data and process!

A. Image Analysis

• Ideal for: Field techs, educators, visual documentation
• Strength: Lightning-speed image understanding
• Model: Gemma 3 (Starter+) / Mistral Small (Pro+)

Try this Prompt:

"Extract all text from this supply chain invoice. Identify vendor, PO number, total due, and delivery date. Flag any discrepancies."
(📷 Upload invoice photo)

B. Creative Ideation

• Ideal for: Marketers, creatives, ideation
• Strength: ChatGPT-level innovation without data leaks
• Model: GPT-OSS 120B (Pro+)

Try this Prompt:

"As our Chief Innovator: Propose 3 viral campaigns for eco-headphones. Include Gen-Z hashtags and TikTok hooks."

C. Complex Analysis

• Ideal for: Researchers, compliance officers
• Strength: Dissects complex documents like a pro
• Model: DeepSeek R1 (Pro+)

Try this Prompt:

"Cross-check clinical trial data against FDA guidelines. Output risk matrix with confidence scores. Provide as a text version and as JSON output. [paste clinical trial data]"

D. Multilingual & Technical Work

• Ideal for: Developers, international teams
• Strength: Fluent in 100+ languages + code
• Model: Qwen3 Coder (Pro+)

Try this Prompt:

"Debug this Python API snippet. Explain errors in English and Japanese for Tokyo engineers. Add optimization comments. [paste snippet]"

E. Therapy Session Productivity

• Ideal for: Daily productivity, therapists
• Strength: Balances reasoning power with accessibility
• Model: Llama 3.3 (All plans)

Try this Prompt:

"Convert session notes into SOAP format:
[Patient] expressed anxiety about job transition
Discussed grounding techniques
Assigned thought journaling
———
Highlight themes and progress markers."

F. Organizational Efficiency (Teams)

• Ideal for: HR/People Teams, Market Research, Data Analysts
• Strength: Complex analysis, advanced reasoning
• Models: DeepSeek R1 671B (Pro+)
• Try this Prompt:

"Redline HR policy draft for GDPR/California law conflicts. Highlight high-risk sections in red. [paste draft]"

G. Personal Life Enhancement

• Ideal for: Personal Life, Family, Friends
• Strength: Conversations, Basic Research
• Models: Gemma 3 (Starter+), Mistral Small 3.1 (Pro+), or Llama 3.3 (All plans)
• Try these Prompts:

Family Trip

"Plan a budget-friendly weekend trip from Seattle:
- 2 adults, 1 child (age 7)
- Educational activities
- Max $600 total"

Happy Hour Debate

"Settle a debate: Explain quantum entanglement using baking analogies"

Apology Text

"Draft a thoughtful apology message to my sister (tone: warm but sincere). Here is what happened: [explain yourself]"

H. Daily Productivity

• Ideal for: Volunteering, Personal Life
• Strength: Communications, Conversations
• Models: Llama 3.3 (All plans)
• Try this Prompt:

Draft a polite email to parents in my child's class about the holiday party. Tone: friendly but professional. Here are the notes from our last planning meeting: [paste notes]"

Customize the Model and Personality

You can tell the model how you want it to function, communicate, and behave. This can go by various names: system prompt, character, and instructions. Point is, the power is yours to personalize the model to your needs.

Tap the 🤖 icon above the text box when starting a new chat. Anything you put in this box will apply to the entire conversation in that chat. Set system prompts like "Respond as a skeptical scientist", "You're a sarcastic historian", "Use medical terminology", or "Use LEGO analogies for complex topics."

Using Maple Like a Pro

1. Switch models: Tap current model name → Choose new assistant.
2. Analyze images: Select Gemma/Mistral → Tap 📷 → Upload.
3. Supercharge with prompts: Click 🤖 to add "Format as legal memo" or "Tone: supportive therapist".

FAQs

Q: Why can't Free users access image models?

Image processing requires more processing power. Starter+ plans help us cover these costs while keeping privacy intact.

Q: Is GPT-OSS 120B really like ChatGPT?

Yes—GPT creative and analytical power, but zero data retention. Your marketing brainstorms stay yours.

Q: How did you write this guide?

We used DeepSeek R1 671B inside Maple to research, draft, and refine every section—encrypted end-to-end.

Q: What are the context windows for each model?

Higher "K" means longer chats.

• Llama 3.3 70B: 128K
• Gemma 3 27B: 128K
• Mistral Small 3.1 24B: 128K
• Qwen 2.5 72B: 128K
• Qwen3 Coder 480B: 128K
• GPT-OSS 120B: 128K
• DeepSeek R1 671B: 128K

Your Privacy-Powered Future Starts Now

Maple proves you don't need to surrender data for world-class AI. Whether you're a therapist protecting client notes or a team auditing sensitive contracts, our models deliver specialized intelligence, wrapped in encryption.

Verify our security live: Maple Attestation Proof
Upgrade or get started: Sign Up Now

Keep chatting.

We started with two plans: Starter and Pro. Honestly, we weren’t even sure if people would pay for a privacy-first AI assistant. But they did.

We’ve been blown away by the response. It’s not just that users care about privacy—they’re getting real, daily utility from Maple. Our Pro plan quickly became the most popular tier, powering everything from confidential legal drafting and therapy notes to startup brainstorming and research workflows. Maple has grown faster and further than we expected.

But some of you? You’re on another level. You’re building businesses on Maple. Running professional workflows. Automating high-stakes work. For you, Pro isn’t enough.

Introducing the Max Plan

20x the power of Pro. $100/month. Unlimited ambition.

Max was built for people who need Maple without limits—more speed, more usage, more support, and more early access to what’s coming next. Whether you’re processing sensitive documents, running long-form research chats, or just tired of hitting caps, Max gives you the freedom to work without friction.

What’s in the Max Plan?

• 🔒 All the privacy features of Maple Pro
• ⚡️ 20x more usage than Pro
• 🛠 Priority support
• 🚀 Early access to new features & models
• 🖼 Image Upload
• 📄 Document Upload
• 🧠 Access to top-tier models: DeepSeek R1, Qwen 2.5, Gemma 3, Mistral Small 3.1

All this with privacy you can prove yourself.

Goodbye, Starter. (Sort of.)

As of today, the Starter plan is closed to new users.

To our early adopters on Starter: thank you. You believed in private AI before it was cool. You’ll be able to stay on Starter as long as you're subscribed. Yes, there may come a time when it sunsets—but we’ll cross that bridge when we get there.

What’s Next

We’re just getting started. Here's a peek at what we're researching:

• 📄 Encrypted document RAG (retrieval-augmented generation)
• ☁️ Cloud storage integrations for file uploads
• 🔐 Private live web search
• 🤖 Custom GPTs you control
• 👥 Secure team collaboration & shared history
• 🧾 Role-based permissions for organizations

We're building Maple not just for individuals, but for teams, companies, and entire industries that require confidentiality and performance.

If you're an organization that deals with sensitive data—healthcare, law, finance, journalism—you don’t have to trust us. You can verify us.

Maple runs on open source code and confidential computing. That means you (or your security team) can audit the code, verify the attestation, and see exactly how your data is encrypted and processed—without trusting a black box. No vague promises. Just cryptographic proof.

Upgrade to Max

If you're pushing the limits of Pro, Max is ready for you. Whether you're a founder, researcher, developer, executive, or just someone who talks to AI all day—Max is your new best plan.

Upgrade to Max

Because your data is your data. Not ours.

Whether you're a law firm drafting confidential contracts, a therapy practice analyzing client notes, or a fintech team processing financial data, Maple Teams lets your entire group harness AI without compromising security.

The Privacy Challenge

Many organizations ban or restrict third-party AI tools because they risk exposing proprietary or personal data. Traditional chatbots often store conversations for training, leaving your information vulnerable to breaches, insiders, or even the providers themselves. For organizations in regulated or trust-driven industries, that’s a nonstarter.

Even with restrictions in place, well-meaning employees use AI tools that retain sensitive data so they can be more productive, putting the organization at risk.

That's where Maple stands apart. Built on end-to-end encryption and confidential computing with AWS Nitro Enclaves, Maple ensures that your data remains private to your account. No one—not even our team—can access your plaintext conversations.

We use hardware-level isolation to process queries securely, and our open-source code allows you to verify it yourself.

Key Features for Teams

With Maple Teams, you get all the core features of our Pro plan, plus tools tailored for group use:

• Pooled Usage: Share credits across your team for efficient AI access, whether it's summarizing documents, analyzing images, or running advanced models like DeepSeek R1.
• Simple Onboarding: Invite members easily, manage seats, and track usage from a central dashboard. Team admins can add or remove users, resend invites, and monitor everything without exposing data.
• Early Access Features: Get advanced access to powerful new features in Maple before the rest of the platform.
• Dedicated Support: Priority assistance to keep your workflows running smoothly, including guidance on integrating Maple into regulated environments like healthcare or finance.

Boost Productivity Securely

This isn't just about security—it's about productivity. Therapists can safely brainstorm treatment plans, lawyers can draft compliant documents faster, and accountants can review spreadsheets with AI insights, all while staying compliant with privacy standards.

Our users in these fields already report saving hours each week, and Teams extends that to your whole organization.

Easy Setup

Getting started is straightforward. From our Pricing page, select the Team plan, complete the purchase via Stripe, and invite your colleagues by email.

They'll join with their own Maple accounts, and you'll be chatting securely in minutes. Check out our guide for full details on Team setup and management.

Built for Teams. Built for Trust.

You shouldn’t have to choose between powerful AI tools and strong data governance. With Maple Teams, you can deploy AI across your organization—confident that what happens in your workspace, stays in your workspace.

Upgrade today and experience secure, collaborative AI that scales with your needs. Visit our Pricing page to get started, or reach out to team@opensecret.cloud for a demo.

Your data stays yours—let's put it to work. 🤘

Maple allows any authenticated user to purchase, create, and manage team accounts directly from the application. This guide covers the complete workflow from team purchase to member management.

Setting up your team is 3 simple steps:

1. Purchase the number of seats you need
2. Invite your team members
3. Chat with encrypted AI

Note: Prices might be different than these screenshots

Purchasing a Team Plan

Go to the Pricing page and hit "Start Chatting" on the Team plan.

Log in with your existing Maple account or sign up if you don't have one yet.

You'll be redirected to Stripe checkout where you select the number of seats you need. Minimum is 2. Then pay with your preferred method.
Note: Bitcoin payments are not available for Team accounts.

Managing the Team

After successful payment, you'll be redirected to the team setup flow.

On this screen you can name your team and see the number of available seats.

Invite people to join your team using their email address. You can enter multiple people at one time by putting each email on its own line or separating them with a comma.

Tell your team members to check their email and click the "Accept Invite" button.

Members log in if they already have a Maple account or create one at this step.

Chat with AI

You and your team members are all set to start chatting!

Each user knows they are on the team plan by looking at the bottom of the history pane. Users can see their team name, credit usage, and more.

This is where Admin users access the Team Dashboard and the Subscription. If you want to add more team members in the future, you do it on the Manage Subscription screen.

Additional Management Options

Edit Team Name
Click the pencil icon next to team name to change it.

Resend an Invite
Invites expire after a few days, so you might need to send a new one. Go to the Team Dashboard and hit "Resend Invite" next to that user.

Remove a User
If a person is no longer part of your organization, you can go into the Team Dashboard and click the ⠇button next to their email, and choose "Remove from Team".

The user is removed from your team but their Maple account is still active. They revert to being a Free account.

Getting Help

If something isn't working right for you, reach out to us by email support@opensecret.cloud or hop into our Discord.

Onboard your organization to Maple and start securing your important data today!

This feature opens up new dimensions of productivity, efficiency, and privacy-first innovation for professionals and individuals alike. Whether you're reviewing a contract, translating a sign, or analyzing your finances, Maple AI now lets you bring documents and images into the conversation—without compromising your data security.

Maple is now smarter than ever—and ready to streamline your workflow.

What’s New: Document and Image Upload

With this new release, Maple AI supports two major upload features:

• Image Upload: Snap a photo or choose one from your camera roll—Maple can now interpret and analyze images on the fly.
• Document Upload: Upload files in formats like PDF, DOCX, Excel, and more for Maple to analyze, summarize, or explain. Available on Desktop, iOS, and Android apps. Web support coming soon.

This unlocks a variety of use cases:

• Extract meeting notes from a presentation deck
• Proofread and improve a school assignment
• Scan and interpret a handwritten note or recipe
• Summarize a lengthy essay or article
• Translate a printed menu or sign while traveling

All with end-to-end encryption and the privacy-first design Maple users trust.

Are my files safe and secure?

Yes. File upload for both images and documents follows the same secure process as all of our data processing. We use encryption to securely transfer the file. Once in the cloud, secure enclaves handle the file parsing and pass the data to the AI that also uses secure enclaves.

Your data is protected at every step in the process.

Image Upload: Unlocking Visual Possibilities

Visual tasks just got easier and more secure with Maple.

🌍 Travel and Translation

Lost in translation? Snap a photo of a menu, sign, or product label while abroad, and let Maple instantly translate or interpret it. No need to type or search manually—just point, shoot, and understand.

🪴 Nature Exploration

On your next hike, take a photo of an unfamiliar plant and ask Maple what it is. Whether you're a hobbyist botanist or simply curious, Maple gives instant feedback and species information—with zero data leaving your encrypted session.

Document Upload: Smarter Analysis for Better Results

Text-based tasks now benefit from intelligent AI analysis, all inside your secure workspace.

🧠 Academic & Writing Help

Upload your school paper or essay draft and let Maple review it for grammar, clarity, tone, and flow. Or drop in a lengthy article and get a well-organized summary in seconds.

📊 Work Smarter

Uploading a slide deck? Maple will extract key points, structure them as meeting notes, and help you prep recaps or follow-ups fast—saving you precious time.

Available on Pro and Teams

These powerful tools are available to Pro and Team plan members. If you're on Free or Starter, consider upgrading to get the most out of your confidential AI.

What’s Next

We’re just getting started. Coming soon:

• Upload multiple files in a single prompt
• Expanded support for more file types
• Integration with cloud services, possibly Google Drive and others
• Shared document access and team collaboration tools for businesses

Upgrade Today

Upgrading to Pro or Team gives you access to Maple's most powerful features—without compromising on privacy. Whether you're working solo or collaborating with a team, Document and Image Upload will save you time and elevate your decision-making.

🚀 Upgrade Now to Access Document and Image Upload

In Summary

This new feature expands what’s possible in private AI, reaffirming our commitment to secure, encrypted productivity. With Document and Image Upload, Maple AI becomes an even more indispensable tool for your work and life.

We’d love to hear how you’re using the new features—share your experience on social media and tag @TryMapleAI, if it's something you can share 😀

Your productivity just got a major upgrade. Start exploring the possibilities with documents and images in Maple today.

Technical Specs

• Supported File Types
  • Images: .jpg, .png, .webp
  • Documents: .pdf, .doc, .docx, .txt, .rtf, .xlsx, .xls, .pptx, .ppt, .md
• File Limitations
  • Max file size: 5MB
  • One file per reply

About Maple AI

Maple AI is the world’s most private AI chat assistant, built with end-to-end encryption, confidential computing, and zero data mining. Whether you’re a privacy-conscious individual, an independent professional, or an enterprise in a sensitive industry, Maple delivers the power of AI without compromising your trust.

DeepSeek R1 demonstrates its versatility by excelling in a wide range of diverse professional contexts. Whether supporting mental health professionals, assisting legal experts, or empowering creative strategists and activists, the model provides advanced capabilities tailored to specific workflows. Here are a few examples of how DeepSeek R1 enhances productivity across different fields.

DeepSeek R1 in Supporting Therapists

DeepSeek R1 excels in analyzing client data, providing insights that support therapists in diagnosis and treatment planning. It assists professionals in understanding complex client behaviors, enabling them to identify patterns and potential therapeutic approaches with greater accuracy. By offering data-driven suggestions, DeepSeek R1 enhances a therapist's ability to create personalized treatment plans tailored to each client's needs. This support allows therapists to focus more on the human aspect of care, knowing they have reliable analytical backing.

As always, you, the therapist, are in control of the treatment. Check out our post on recommended ways to use AI safely with therapy.

DeepSeek R1 in Legal Drafting

DeepSeek R1 understands nuanced legal terminology, enabling the efficient drafting of accurate and compliant documents. It helps lawyers stay updated with legal changes, ensuring documentation reflects current regulations. By automating routine tasks and providing tailored language, DeepSeek R1 streamlines legal workflows, allowing professionals to focus on high-value tasks. Its comprehension of complex contexts ensures the creation of precise and relevant documents, reducing the need for extensive research.

An advanced model like DeepSeek is excellent to have in Maple's confidential AI, as many legal and regulatory bodies discourage or prohibit the use of tools like ChatGPT due to concerns over data privacy and attorney-client privilege. Maple encrypts your data so only the user and the AI can see it.

DeepSeek R1 in Creative Outputs

DeepSeek R1 excels in generating innovative content based on the input and data provided by users, aiding in the creation of engaging marketing materials. It supports activists by helping them craft clear and compelling messages, and it organizes data provided by users to advocate more effectively. DeepSeek R1 helps identify effective communication strategies, amplifying their impact in promoting social, environmental, or political issues.

The version of DeepSeek within Maple is an open-source model and is in no way affiliated with the creators of DeepSeek. Data is not shared with any government or used for training purposes. It stays private and confidential in your account.

Security with Tinfoil

Your security is paramount. DeepSeek R1 is hosted through Tinfoil, an innovative provider of confidential computing, ensuring that all data processing is handled securely and meets high standards of privacy.

Why Upgrade?

DeepSeek R1 is the latest feature designed to support professionals and organizations. Some features may become available to all users over time, but others will remain exclusive to our Pro and Teams tiers. Upgrading assures you will have access to the best confidential AI features.

Conclusion

At Maple AI, privacy is at our core. With DeepSeek R1 and our partnership with Tinfoil, we continue to enhance our secure AI computing offerings. While DeepSeek R1 is currently available only to Pro and Teams users, we remain committed to providing valuable updates to all our users. Many more things are cooking!

P.S. You can thank R1 for helping me write this post 🤘 Give it a try yourself!

Privacy First: Your Conversations, Secured

At Maple AI, we believe privacy isn't a feature—it's foundational. Apple's commitment to privacy aligns perfectly with our mission, making your iPhone and iPad ideal companions for Maple AI. With end-to-end encryption, confidential computing, and open-source transparency, your conversations remain exclusively yours—private, secure, and accessible anytime, anywhere.

Designed for Busy Professionals and Prosumers

Whether you're a therapist managing sensitive client information, a lawyer drafting confidential documents, or an entrepreneur brainstorming your next big idea, Maple AI delivers secure productivity. Our app helps you:

• Securely Collaborate: Draft confidential documents or proposals on-the-go.
• Brainstorm Privately: Generate creative ideas and outlines securely from anywhere.
• Seamlessly Sync: Move effortlessly between your desktop, iPhone, and iPad, always picking up right where you left off.

Privacy for Everyday Life

For individuals who prioritize privacy in their daily routines, Maple AI offers a safe space to handle personal tasks securely. Whether you're organizing your family schedule, journaling private thoughts, or simply chatting privately, Maple AI keeps your personal information confidential and secure.

Teams: Take Your Secure AI Work Mobile

For larger teams and organizations, Maple AI extends "on-premise" privacy into the cloud and onto your mobile devices. Your entire team can now securely access AI-powered assistance wherever their work takes them, maintaining complete data privacy and control.

Pricing

Maple AI offers flexible pricing plans to suit everyone. Our Free plan allows users to explore core features without any cost. Individuals seeking more robust usage can upgrade to the Starter or Pro plans for expanded capabilities. For larger teams and organizations, our Team plan provides pooled usage, enhanced collaboration, and dedicated support.

Check out our Pricing page to learn more and get started!

Experience Private AI Anywhere

Download Maple AI from the App Store today and experience truly private AI productivity on your iPhone and iPad.

Your privacy matters—let Maple AI keep it secure.

In a 2025 survey by the American Psychological Association, 56% of psychologists said they had used AI tools for work at least once in the past 12 months, up from 29% a year earlier, and 67% were concerned about data breaches.¹ MIT Technology Review has reported on therapists using ChatGPT without telling clients, including a client who watched his therapist put what he was saying into it during a session.²

Can therapists use ChatGPT or other AI tools?

Yes, for work that contains no client information, and for information that is properly de-identified under 45 CFR 164.514(b). For identifiable client information, HHS guidance says a HIPAA-covered practice that uses a vendor to process or store it without a business associate agreement "is in violation of the HIPAA Rules."³ State law and your ethics code apply too.

On consumer plans of ChatGPT, Claude and similar tools, what you type reaches the provider in a form the provider can read. The provider's policies decide who can review it, how long it is kept and whether it trains future models.

The diagram below shows where what you type can be read, on a consumer plan and on Maple.

Review can mean people, and it can reach the police. Anthropic's consumer terms reserve the right to report your inputs and outputs to law enforcement "at our sole discretion." OpenAI has said chats about harming others go to a human review team that can refer an imminent threat to police, though it said in 2025 it was not referring self-harm cases.

Session material is full of what those systems look for: a client's threat, a disclosure of abuse, suicidal thoughts. Paste it into a personal account and a stranger may read your client's words. We cover the cases so far in Can AI Companies Report You to the Police?

Today's policy is also only today's. A company can rewrite its terms whenever it chooses, and the chats it already holds are still there when it does. OpenAI's choice not to refer self-harm cases is a policy, not a limit on what it can do, and Anthropic announced a usage policy update as recently as October 2026. Large platforms answer to regulators, advertisers and public pressure, and their rules tend to move with it.

Consumer plans also come without a BAA. OpenAI lists the products it offers with a BAA, such as ChatGPT for Healthcare and ChatGPT for Clinicians, and Anthropic offers one for its API and Enterprise plans. The free and personal plans most people sign up for are not on OpenAI's list, and Anthropic's BAA page covers its commercial products (checked 10 October 2026). The test for any tool is whether a signed BAA covers your account.

Every AI task falls into one of three tiers.

What can you do with AI and no client information?

Most of the writing that fills a clinician's evenings contains no client information at all: handouts, form language, note templates, letters, policies and website copy. None of it involves protected health information, so HIPAA's vendor rules don't come into play. Your own review still does.

These four prompts need no client detail.

You can also hand the assistant a document with no client in it, such as a journal article or a blank intake packet. In Maple that works through document upload, and our getting-started guide covers setup.

Is removing the client's name enough to de-identify a note?

No, removing the name is not enough. Under HIPAA's Safe Harbor method, 45 CFR 164.514(b)(2), you must remove 18 kinds of identifiers and have no actual knowledge that what remains could identify the person.⁴ HHS guidance says a data set that contained patient initials "would not meet the requirement of the Safe Harbor method for de-identification."⁵

The identifiers include any geographic unit smaller than a state, every date element except the year, phone numbers, email addresses and record numbers. They also cover the client's relatives, employers and household members, so a sister's name or the name of a workplace counts.

If you need a label, the rule allows a code that "is not derived from or related to information about the individual." A random "Client 7" qualifies. A code built from initials or a birth date does not.

Why can a detailed story still identify a client?

HHS says a covered entity has actual knowledge "if it concludes that the remaining information could be used to identify the individual." Therapy material is full of that kind of detail: an unusual job, a publicized event, a recognizable family situation. Its guidance adds that clinical narratives that document a patient's history or lifestyle "are information rich and may provide context that readily allows for patient identification."

So de-identification fails more often than people expect. A practical rule: strip the 18 identifiers, then reread the story and ask whether someone who knows the client would recognize them. If the answer is maybe, generalize further or leave it out.

Here the design of the tool starts to matter. If a detail slips through on a consumer plan, a readable copy sits with the provider, where it can be scanned, reviewed by staff and, under some terms, reported. In a tool built so that the company running it cannot read your chats, the same slip is not readable by that company's staff. Treat that as a safety net under careful de-identification, never as a reason to skip it.

Does encryption make an AI tool HIPAA compliant?

No, encryption alone does not. HHS's cloud computing guidance says a vendor that stores or processes a covered entity's electronic protected health information (ePHI) is a business associate "even if the entity cannot actually view the ePHI."³ The practice and the vendor still need a business associate agreement, however strong the encryption.

HHS settled the question in one sentence about cloud service providers (CSPs) that hold encrypted data and no key.

A BAA is the contract in which a vendor takes on HIPAA duties for the client information it handles for you. Under 45 CFR 160.103, a business associate is a person who "creates, receives, maintains, or transmits protected health information" on a covered entity's behalf.⁶ The test is what the vendor does with the data, whether or not its staff can read it.

Encryption still earns its place by lowering breach risk, but HHS says "such protections alone cannot adequately safeguard" the data as the Security Rule requires. Under today's Security Rule, encryption is an "addressable" safeguard.⁷ HHS proposed in January 2025 to make it mandatory, and that proposal was not final when we checked on 10 October 2026.⁸

No AI tool is government-certified, either. The same guidance says the Office for Civil Rights "does not endorse, certify, or recommend specific technology or products." When a vendor says "HIPAA compliant," ask what stands behind the phrase, starting with a BAA.

A BAA makes a tool legally appropriate for client information. It doesn't change the technology: the vendor's systems still handle your text in readable form, and the protection is the vendor's promise and the law behind it.

Does HIPAA apply to private-pay therapists?

Not always. HIPAA covers a health care provider "who transmits any health information in electronic form in connection with a transaction covered by this subchapter," such as an electronic insurance claim (45 CFR 160.103). HHS says an entity that is neither a covered entity nor a business associate "does not have to comply with the HIPAA Rules."⁹

CMS says providers "who submit HIPAA transactions, like claims, electronically are covered," and standard transactions are not limited to claims. If you or a billing service acting for you ever send them, you may be covered for your whole practice. CMS publishes a covered entity decision tool.

Being outside HIPAA does not make a consumer chatbot a safe place for session notes. State confidentiality law, the state AI laws below and your licensing board's ethics code still apply.

Do therapists need client consent to use AI?

For AI that touches a client's care, the professional bodies say yes. APA says psychologists "have an ethical obligation to obtain informed consent by clearly communicating the purpose, application, and potential benefits and risks of relevant AI tools."¹⁰ Illinois, Colorado and Maine have also passed laws requiring client consent when AI is used with a recorded or transcribed session.

The counseling bodies agree:

• ACA: "Counselors should clearly inform clients about the use of AI tools in their counseling process, explaining their purpose and potential benefits."¹¹
• NBCC: "Obtain informed consent specifically regarding AI use" and "Allow client choice to opt out of AI tools."¹²

None of these documents exempts back-office drafting with no client information, so a short statement in your intake paperwork about how you use AI is the cautious choice. A licensed clinical social worker explains what is at stake:

This confidentiality allows clients to be vulnerable and honest without fear of their information being misused.

Which states require consent for AI in therapy sessions?

This table has only the states whose legislature pages we read ourselves, checked 10 October 2026.

Other states regulate disclosure instead of consent. Louisiana, for example, requires a licensed healthcare professional to "verbally disclose" a recording that will be transcribed by AI before it starts.¹⁶ Check your own state and licensing board, because these rules are changing.

What should you ask an AI vendor?

Ask every AI vendor, us included, six questions about the plan you would use: whether a BAA covers it, who can read what you type, whether your text trains models, what is kept, how deletion works and which other companies receive your data. Get the answers in writing.

1. Will you sign a BAA that covers my account and plan?
2. Who at your company can read what I type, and is that a policy or a technical limit?
3. Is my text used to train models?
4. What do you keep, including logs and backups, and for how long?
5. Can I delete it, and what does deleting do?
6. Which model providers and cloud hosts receive my data?

How does Maple handle what you type?

Maple encrypts your chats on your device and decrypts them only inside a secure enclave running open-source code you can verify. No one at Maple can read your chats. We don't offer a BAA today, and getting there is a work in progress, so for HIPAA-covered clinicians Maple is a tool for tier 1 work and properly de-identified tier 2 work.

Maple can't see what you send to the models or what they send back, and no third party can read your chats. Your encryption key is managed for you inside the secure enclave.

You can check the design yourself: our proof page shows a live attestation you can compare against the code we publish. Here are our answers to the six questions.

We don't offer a BAA today. We're working toward a BAA and a SOC 2 audit, and we'll say so here when that changes.

Until then, HIPAA-covered clinicians should not enter identifiable client information into Maple. We couldn't see it if you did, and no third party can read your chats. The law still asks for a signed agreement, and HHS says being unable to view the data does not exempt a vendor from that.

What Maple offers today is tier 1 and tier 2 work with a stronger backstop. If your de-identification misses something, the slip lands where our staff can't read it, and there is nothing for us to review, flag or hand over. Maple is a general-purpose assistant, not a clinical note-taking service, and our privacy notice covers the account data in full.

What should you check before you paste anything into AI?

Decide which tier the text belongs to before it goes into the chat box, because that is the step you fully control. Afterward, the output is yours. APA says psychologists "remain responsible for final decisions and must not blindly rely on AI-generated recommendations."¹⁰

1. Tier 1: go ahead in a tool you trust.
2. Tier 2: confirm all 18 identifiers are gone and that nobody who knows the client could recognize the story.
3. Tier 3: stop unless a BAA covers your account, the client has been told or has consented, and your state's rules are met.
4. Afterward: check facts and tone, move what you keep into your own files, and delete chats you no longer need.

Most of the time AI can save you sits in tier 1. Start there with one handout or one template.

Sources

1. American Psychological Association, Among psychologists, AI use is up, but so are concerns, December 9, 2025. Practitioner Pulse Survey of 1,742 psychologists, fielded September 2025.
2. MIT Technology Review, Therapists are secretly using ChatGPT. Clients are triggered., September 2, 2025. Reporting on therapists' undisclosed use of ChatGPT.
3. US Department of Health and Human Services, Guidance on HIPAA and Cloud Computing, content last reviewed December 23, 2022; checked 10 October 2026. Business associate status without a key, the BAA requirement, the limits of encryption, and that OCR certifies no products.
4. Code of Federal Regulations, 45 CFR 164.514: De-identification of protected health information, checked 10 October 2026. Safe Harbor, Expert Determination and re-identification codes.
5. US Department of Health and Human Services, Guidance Regarding Methods for De-identification of Protected Health Information, content last reviewed February 3, 2025. Initials, actual knowledge and free-text narratives.
6. Code of Federal Regulations, 45 CFR 160.103: Definitions, checked 10 October 2026. Business associate and covered entity.
7. Code of Federal Regulations, 45 CFR 164.312: Technical safeguards, checked 10 October 2026. Encryption as an addressable specification.
8. Office of Information and Regulatory Affairs, HIPAA Security Rule to Strengthen the Cybersecurity of Electronic Protected Health Information, RIN 0945-AA22, checked 10 October 2026. Proposed rule published January 6, 2025; listed as a long-term action.
9. US Department of Health and Human Services, Covered Entities and Business Associates, content last reviewed August 21, 2024. Who must comply with the HIPAA Rules.
10. American Psychological Association, Ethical guidance for AI in the professional practice of health service psychology, created June 2025, updated December 2025. Informed consent and responsibility for decisions.
11. American Counseling Association, Recommendations for Practicing Counselors and Their Use of AI, checked 10 October 2026. Transparency and informed consent.
12. National Board for Certified Counselors, Ethical Principles for Artificial Intelligence in Counseling, April 12, 2024. Consent, opt-out and counselor responsibility.
13. Illinois General Assembly, Public Act 104-0054: Wellness and Oversight for Psychological Resources Act, effective August 1, 2025.
14. Colorado General Assembly, HB26-1195: Psychotherapy Artificial Intelligence Restrictions, signed June 3, 2026. Official bill summary.
15. Maine Legislature, LD 2082, Public Law 2026, chapter 687, approved April 13, 2026.
16. Louisiana State Legislature, HB 475, Act No. 649, effective August 1, 2026. Verbal disclosure before AI transcription.
17. Maple, Proof: how Maple's encryption can be verified, checked 10 October 2026.
18. Maple, Privacy notice, checked 10 October 2026.

Frequently Asked Questions

Is Maple HIPAA compliant?

HIPAA has no certification, and HHS does not certify or endorse products, so the label means little on its own. What matters for a covered practice is a business associate agreement (BAA). Maple does not offer a BAA today, and we are working toward one. Until a BAA covers your account, use Maple for work with no client information or with properly de-identified information. No one at Maple can read your chats, so the limit is legal, not technical: the law asks for a signed agreement, and HHS says a vendor that cannot view the data still needs one.

Can I put client notes into ChatGPT?

Not on a consumer plan if you are a HIPAA covered entity. HHS guidance says a covered practice that uses a vendor to process or store electronic protected health information without a business associate agreement is in violation of the HIPAA Rules. OpenAI offers a BAA for specific business and healthcare products, so the test is whether a signed BAA covers your account. HIPAA aside, a consumer provider can read what you type, flagged chats can go to human reviewers, and some terms allow reports to law enforcement. A blank note template with no client information is fine in any tool you trust.

Are initials enough to de-identify a note?

No. HHS guidance says a data set that contained patient initials would not meet the requirement of the Safe Harbor method for de-identification. Safe Harbor requires removing 18 kinds of identifiers, including those of relatives, employers and household members, and having no actual knowledge that what remains could identify the person. A random code such as Client 7 is allowed if it is not derived from information about the client.

Do I need client consent to use AI in my practice?

For AI that touches a client's care, APA, ACA and NBCC all call for disclosure and informed consent. Illinois, Colorado and Maine have passed laws requiring client consent when AI is used with a recorded or transcribed session, as of 10 October 2026. None of the professional guidance exempts back-office drafting with no client information, so a short statement in your intake paperwork is the cautious choice. Check your own state and licensing board, because these rules are changing.

Does Maple train on what I type?

No. We do not train on your conversations, and we could not: no one at Maple can read your chats. They are encrypted on your device and decrypted only inside a hardware-isolated enclave running open-source code you can verify. Your chat history is stored encrypted.

Does HIPAA apply if I only take private-pay clients?

Not always. HIPAA covers a health care provider who transmits health information electronically in connection with a standard transaction, such as an electronic insurance claim, and HHS says an entity that is neither a covered entity nor a business associate does not have to comply with the HIPAA Rules. State confidentiality law, state AI laws and your ethics code still apply. CMS publishes a decision tool for close calls.

Why Most Planning Systems Stall

Staring at a messy to-do list can freeze progress before the day even starts. We've all been there. We know there is an endless list of things to accomplish, but we struggle with where to begin. Two universal bottlenecks derail otherwise solid productivity routines.

Cognitive overload
The brain can juggle only a handful of open loops before decision-making slows.

Context switching
Each “What should I tackle next?” moment burns time and willpower.

Traditional fixes (bullet journals, sticky notes, kanban boards) help, but they still depend on manual discipline. Miss a day and the paper backlog mushrooms overnight. Executive assistants can handle the organization but are expensive to hire and train up.

Your personal AI executive assistant has entered the chat.

Maple AI’s daily-planning workflow offloads the tedious parts—capture, prioritization, scheduling—while keeping you in control of the decisions. Because Maple encrypts everything end-to-end, you can feed it sensitive details (client names, project budgets, personal errands) without worrying about who’s looking. The result is a living plan that adapts as reality shifts.

Your personal AI assistant takes three important steps and walks you through them in one easy process.

Would you rather watch how this process works? Check out our video walkthrough.

Three Pillars Behind an Effective Schedule

Before jumping into the tutorial, let’s unpack the science each step relies on:

1. Externalization
   Theory: Getting thoughts out of your head reduces working-memory load (Sweller’s Cognitive Load Theory) and cuts stress signals in the brain.
   Practice: Speaking stream-of-consciousness into Maple captures more nuance than typing, and lets you think in “raw form” instead of polished sentences.
2. The Eisenhower Matrix
   Theory: Introduced by U.S. President Dwight D. Eisenhower, the matrix separates tasks by importance (“Will this advance my goals?”) and urgency (“Does delaying have near-term consequences?”). Decision-science research shows that ranking on two distinct axes beats simple to-do lists for long-term achievement.
   Practice: Maple scores each axis from 1–3, sums the numbers, and sorts. A smaller total means “do first.” This numeric tweak turns a conceptual matrix into a concrete, AI-friendly algorithm.
3. The Pomodoro Technique
   Theory: 25-minute “sprints” exploit the brain’s natural attention span (the vigilance decrement curve) while preserving flow. Frequent micro-breaks replenish dopamine, and a longer pause every four cycles prevents fatigue (Pashler, 2001).
   Practice: Maple converts priorities into Pomodoros, inserts 5-minute resets, and even groups small tasks so the timer always feels meaningful.

The Maple AI Workflow (Deep Dive)

Step 1 – Capture Everything in One Go

1. Open Maple on iOS, macOS, Linux, or the web. Get your free account at maple.ai.
2. Tap the microphone icon.
3. Talk freely—client projects, deadlines, chores, “remember to call the dentist,” even abstract worries.

End your brain-dump with:

“Please turn this into a concise to-do list.”

Why this works
Voice entry taps your natural pace (~140 words per minute), often doubling what you’d type. Because Maple encrypts on your device, you can mention NDA-protected projects or family health appointments without filtering yourself—a huge win for completeness.

Pro-tip: If you’re shy about speaking aloud in a cubicle, plug in earbuds with an inline mic and talk at normal volume. Maple’s speech model handles low-volume whispers well.

Step 2 – Rank Tasks With the Eisenhower Matrix

Paste—or say—a follow-up:

“Rank each task by importance and urgency using the Eisenhower matrix. Add the two numbers, sort by priority, and return a single list.”

Under the hood Maple intelligently categorizes the type of task, estimates importance and urgency, calculates a score, and combines work + personal items into one ladder. That prevents the classic trap of an “all-work” list that ignores picking up prescriptions—or an “all-home” list that torpedoes tomorrow’s client pitch.

Fine-tuning on the fly
If Maple misreads your priorities, just converse:

• “Move ‘draft proposal’ above ‘lunch with mentor.’”
• “Merge ‘buy groceries’ and ‘pick up medication’ into one errand block.”

The model instantly recalculates ranks without forcing you to re-score everything.

Step 3 – Build a Schedule With Pomodoro Blocks

Tell Maple:

“My work window is 9:30 am – 4:30 pm. Turn the ranked list into 25-minute Pomodoros with 5-minute breaks after each, a 45-minute lunch, and group any tasks that take <15 minutes.”

What Maple returns

Adapting in real life
Unexpected meeting pops up? Tell Maple, “Shift everything after 1 pm forward 30 minutes,” or “Defer the lowest-priority Pomodoro to tomorrow.” Changes remit in seconds.

Reuse the System Every Morning

After your schedule looks right, ask:

“Create a reusable prompt that captures steps 1–3 so I can paste it tomorrow.”

Save that block to your notes app of choice. Each new day:

1. Start a fresh chat (keeps token use low)
2. Paste the prompt
3. Voice-dump
4. Approve the ranked list
5. Receive a tailored timetable

The whole cycle averages 5-10 minutes—less time than scrolling social media while coffee brews. (We all scroll longer than we expected.)

Daily Planning Prompt You Can Use

Finally, here is what you've been waiting for. This is the full prompt I have developed over the last few months. Feel free to use it, change it to meet your needs, and share it with others.

Daily Priorities with Eisenhower Matrix and Pomodoro Technique

You are my executive assistant.

====
TASK 1 – INGEST MY DAILY TASK RAMBLINGS AND PRIORITIZE
====
I want you to listen to all of my ramblings about the upcoming day, the things that are on my mind that I would like to try and get done today. Please listen to my ramblings and create a concise to-do list with Importance (I), Urgency (U), and Priority (P) scores for each task. The task list should be concise, one line per task, without secondary information. I want to get the most important and highest value items done first. Prioritize my tasks using the Eisenhower Matrix framework, where Importance (I) and Urgency (U) scores range from 1 (highest) to 3 (lowest). Calculate the Priority (P) score by adding the I and U scores, and sort the tasks by P score, with lower scores indicating higher priority. Consider both the impact and deadlines of each task when assigning I and U scores. If I tell you that an item has a specific importance or urgency, trust my value. If I tell you about any scheduled meetings with times during this task, don’t put them into the prioritized list, but put them in a separate list that will be used in Task 3.

====
TASK 2 – BACK AND FORTH CONVERSATION TO ADJUST PRIORITIES
====
After you give me the initial prioritized list, I will review. I'd like to have a back-and-forth conversation to adjust the priorities and make sure everything is in the right order. Wait for me feedback before proceeding to the third task. Remember, you are my executive assistant, so you will have a concise and productive conversation with me to make sure I am working on the most important things first.

====
TASK 3 – POMODORO GENERATION
====
Once I'm satisfied with the priority list, please wait for my signal (e.g., 'Generate Pomodoros' or 'Create Schedule') before generating a suggested Pomodoro schedule for me. To create the schedule, please ask me for my available work hours for the day (e.g., "What are your work hours today, from X to Y?") and if I have any pre-scheduled meetings or time commitments that you should take into account when planning my Pomodoros. It I told you any scheduled meetings during Task 1, this is where you can pull those in. Assume a standard Pomodoro duration of 25 minutes, with 5-minute breaks in between, and a longer break every 4-6 Pomodoros. Take into account the estimated time required for each task and group similar tasks together when possible. Provide a breakdown of the suggested Pomodoro schedule, including the tasks to be worked on during each Pomodoro and any recommended breaks.

Let's get started! Here is my rambling:

Security Matters: Why Encryption Isn’t Optional

Other AI planners store raw text on their servers. That means:

• App company staff can read your workload.
• Client confidentiality clauses may be violated.
• Data breaches can spill your intimate details onto the web for sale.

Maple’s design flips the default:

1. Device-side encryption before transit
2. Secure enclaves to handle keys—no plain-text ever lands in our cloud
3. Zero-knowledge architecture: not even Maple engineers can decrypt your backlog

If you handle legal briefs, client records, or competitive roadmaps, that difference isn’t academic—it’s cryptographic.

Beyond the Basics: Advanced Tweaks

• Switch scoring models. Prefer MoSCoW (Must/Should/Could/Won’t)? Ask Maple to label tasks accordingly and still deliver Pomodoros.
• Integrate calendar data. Paste “Here’s my Google Calendar for tomorrow…” and Maple will avoid existing blocks automatically.
• Team mode. Share a sanitized version of your ranked list with teammates while keeping sensitive notes encrypted in your private chat.

Next Steps

1. Sign up free at maple.ai (10 chats/week on the free tier).
2. Paste the reusable prompt into a new chat.
3. Speak your to do list, approve the ranking, and receive your schedule.
4. Rinse and repeat tomorrow—securely.

A clearer plan isn’t just about finishing tasks; it frees mental bandwidth for strategic thinking, creativity, and (yes) a life outside work.

Give Maple ten minutes, and win back the rest of your day.

Maple AI flips the script: every conversation is end-to-end encrypted so only you can read or reuse your information. In this step-by-step guide you’ll learn how to:

1. See what money details ChatGPT already stores
2. Export those facts with one “Financial Export Prompt”
3. Import them into Maple with a matching “Financial Import Prompt”
4. Ask Maple for hyper-personalized but fully private advice on budgets, raises, retirement, and credit scores

Let’s jump in.

1. What ChatGPT (and other AI services) Remember About Your Money

Every question you’ve asked about budgeting, loans, or investments becomes part of ChatGPT’s long-term memory. That data can resurface in later chats—handy yet risky. Before switching to a privacy-first assistant, you need a clean export.

2. Craft a One-Shot Financial Export Prompt

Open any ChatGPT thread and paste the block below. It instructs the model to search its entire history, pull out personal-finance info only, and package it neatly.

You are an AI language model with full access to:
• The complete text of every message in this conversation
• Any long-term memory, user-profile, or retrieval tools provided by your platform
  (vector store, memory API, knowledge graph, etc.)

========================================================
TASK 1 – PERSONAL-FINANCE EXTRACTION
========================================================
Search *all* available data for explicit or implicit information about **my personal finances only**  
(e.g., salary or income figures, savings targets, personal budgets or spending caps, personal assets or liabilities).

❗ **Exclude** anything tied to a business I own, manage, or invest in—such as corporate budgets, revenue, funding terms, valuations, or ad spend paid from company accounts.

For **each** personal-finance data point you find, output four fields:

  • **Fact** – concise statement of the personal-finance item  
  • **Source** – date/timestamp or message-ID (or “stored-memory” if retrieved from a memory tool)  
  • **Evidence** – direct quotation or close paraphrase supporting the fact  
  • **Context** – one-sentence note explaining how/why that figure was mentioned

Rules:
1. Use only information that is clearly stated or unambiguously implied.  
2. Do **not** invent or guess missing numbers.  
3. Skip any item that is primarily business-related.  
4. If no personal-finance information exists, state “No personal financial information found” and skip Task 2.

Present the results as a bulleted list under the heading **“Personal Financial Facts Found.”**

========================================================
TASK 2 – TRANSFER PROMPT CONSTRUCTION
========================================================
After listing the facts, create a second block labelled **“Transfer-to-Another-LLM Prompt.”**  
This block lets me copy/paste all personal-finance facts into any other model.

Build it exactly like this (but filled in):

### Transfer-to-Another-LLM Prompt
You are an AI assistant that needs to store my current personal-finance facts so future answers remain consistent.  
Please ingest and remember the following JSON payload:

{
  "personal_financial_facts": [
    { "fact": "<Fact 1>", "context": "<Context 1>" },
    { "fact": "<Fact 2>", "context": "<Context 2>" },
    ...
  ]
}

Confirm once the facts are stored in your persistent or session memory.
### End Prompt

Guidelines for the JSON:
• Include **every** Fact from Task 1, but only the *fact* and *context* fields (omit Source/Evidence).  
• Ensure valid JSON (double-quoted strings; escape inner quotes as needed).

Begin now.

After a short wait you’ll receive a structured list of your salary, monthly expenses, retirement balances, and so on. We will call this your Financial Import Prompt. Copy the new prompt, including the data.

Here is an example of one:

Transfer-to-Another-LLM Prompt

You are an AI assistant that needs to store my current personal-finance facts so future answers remain consistent.
Please ingest and remember the following JSON payload:

{
  "personal_financial_facts": [
    { "fact": "Annual salary is $75,000", "context": "Listed in the summary of the user’s personal budget figures recorded on Apr 24 2025." },
    { "fact": "Personal monthly budget totals $6,300", "context": "Overall figure the user provided for regular living costs on Apr 24 2025." },
    { "fact": "401(k) balance is $50,000", "context": "User stated retirement-account value while outlining assets." },
    { "fact": "Brokerage/stock account is worth $15,000", "context": "Declared alongside 401(k) to summarize personal investable assets." },
    { "fact": "Monthly mortgage payment is $1,200", "context": "Part of the housing section in the detailed expense list the user supplied." },
    { "fact": "Monthly property-tax payment is $200", "context": "Also within the housing expense breakout on Apr 24 2025." },
    { "fact": "Monthly homeowner’s insurance is $100", "context": "Final line of the housing sub-category expenses." },
    { "fact": "Monthly utilities total $350", "context": "Sum of electricity, water, gas, and internet given in the utilities block." },
    { "fact": "Monthly electricity bill is $150", "context": "Individual line item inside utilities." },
    { "fact": "Monthly water bill is $50", "context": "Individual utilities entry." },
    { "fact": "Monthly gas (home) bill is $50", "context": "Listed in utilities section." },
    { "fact": "Monthly internet cost is $100", "context": "Last utilities line item." },
    { "fact": "Monthly food spending totals $1,000", "context": "Sum of groceries plus dining-out lines." },
    { "fact": "Monthly grocery spending is $800", "context": "Food sub-item." },
    { "fact": "Monthly dining-out spending is $200", "context": "Food sub-item." },
    { "fact": "Monthly transportation spending totals $600", "context": "Category sum the user supplied." },
    { "fact": "Monthly car loan or lease payment is $300", "context": "Line item within transportation." },
    { "fact": "Monthly vehicle-fuel cost is $100", "context": "Transportation sub-item." },
    { "fact": "Monthly auto-insurance premium is $150", "context": "Transportation sub-item." },
    { "fact": "Monthly vehicle maintenance budget is $50", "context": "Transportation sub-item." },
    { "fact": "Monthly health-insurance premium is $300", "context": "Listed under the dedicated Insurance category." },
    { "fact": "Monthly life-insurance premium is $50", "context": "Same Insurance category block." },
    { "fact": "Monthly debt-repayment total is $300", "context": "Category sum for credit card and personal-loan payments." },
    { "fact": "Monthly credit-card payment is $200", "context": "Debt-repayment sub-item." },
    { "fact": "Monthly personal-loan payment is $100", "context": "Debt-repayment sub-item." },
    { "fact": "Monthly entertainment spending totals $350", "context": "Category sum provided by the user." },
    { "fact": "Monthly movies/events budget is $100", "context": "Entertainment sub-item." },
    { "fact": "Monthly hobbies budget is $50", "context": "Entertainment sub-item." },
    { "fact": "Monthly travel budget is $200", "context": "Entertainment sub-item." },
    { "fact": "Monthly education expenses total $150", "context": "Education category sum in the expense list." },
    { "fact": "Monthly extracurricular-activities cost is $100", "context": "Education sub-item." },
    { "fact": "Monthly school-supplies cost is $50", "context": "Education sub-item." },
    { "fact": "Monthly emergency-fund savings allocation is $417", "context": "Part of the budget breakdown recorded in memory." },
    { "fact": "Monthly miscellaneous spending is $150", "context": "Catch-all category in the user’s budget summary." }
  ]
}
Confirm once the facts are stored in your persistent or session memory.

End Prompt

3. Import Your Data Into Maple AI

1. Start a fresh Maple chat.
2. Maple confirms that it has securely stored the information inside this encrypted conversation.

Paste your Financial Import Prompt that includes your data.

That’s it—the numbers are now locked in a vault even Maple’s engineers can’t open.

4. Put Maple to Work—Privately

Because Maple holds your real figures, you can ask questions generic blog posts never answer:

• “I just received a 3 % raise—how should I tweak my budget?”
• “If I want to retire at 60, what monthly contribution will get me there?”
• “Give me three strategies to lift my credit score within a year.”

Maple crunches the numbers securely and returns guidance tailored to you—no outside eyes involved.

Tip: Long chats consume tokens. When a conversation grows large, run the export prompt inside Maple to get a new Financial Import Prompt, start a new chat with a fresh import, and you’ll keep costs down without losing context.

5. Next Steps & Best Practices

• Keep Maple updated. Whenever your salary, budget, or debt changes, rerun the export/import cycle within Maple so the answers stay current.
• Stay in control. Delete a chat at any time—its encrypted contents become permanently unreadable.

Watch the Walk-Through

The video shows every step in real time, plus live demos of budgeting, retirement planning, and credit-score improvement inside Maple.

Ready to Try It?

1. Run the Financial Export Prompt in ChatGPT (or Grok, Claude, etc.).
2. Paste the resulting data into Maple with the Financial Import Prompt.
3. Ask your most personal money questions—confident they stay personal.

Less tracking, better answers. Welcome to confidential AI.

Why Native Apps Matter

While our web-based version of Maple AI has been well-received, we understand that native apps offer a more immersive and functional experience. By interacting directly with the app, users can enjoy a more streamlined and efficient conversation with our AI assistant. The native app also lays the groundwork for us to introduce more advanced features that are difficult to implement in a web browser, further enhancing the overall user experience.

What to Expect from the Desktop Apps

Our macOS and Linux desktop apps offer the same secure and private AI experience that our users have come to expect from Maple AI. With end-to-end encryption, secure enclaves, and syncing across devices, you can rest assured that your conversations remain confidential and protected. Unique is that you can navigate easily between the app and other apps, improving productivity. We will expand on the unique abilities with these apps in the future. (Looking at you, MCP 👀)

Example Use Cases

Maple AI is designed to be a versatile tool that can be used for both personal and professional purposes. Some examples of how you can use Maple AI include:

• Mental Health Professionals: Therapists, counselors, and psychologists can use Maple AI to generate notes, create treatment plans, and develop personalized therapy strategies while maintaining confidentiality. The AI can help analyze client data, identify patterns, and provide suggestions for evidence-based interventions, all while ensuring that sensitive client information remains protected.
• Researchers and Academics: Researchers working with sensitive data, such as financial information or personal identifiable information, can use Maple AI to analyze and process data while maintaining confidentiality. The AI can help identify patterns, generate hypotheses, and provide insights, all while ensuring that sensitive data remains protected and secure.
• Journalists and Investigators: Investigative journalists and researchers can use Maple AI to analyze large datasets, identify connections, and generate leads while maintaining the confidentiality of sources and sensitive information. The AI can help identify patterns, flag potential leads, and provide suggestions for further investigation, all while keeping sensitive information secure and private.

What's Next?

While we're excited about the launch of our desktop apps for macOS and Linux, we're not stopping there. Our team is hard at work on developing mobile apps for iOS and Android, which will be available soon. We're committed to making Maple AI accessible across all major platforms, ensuring that our users can enjoy a seamless and secure AI experience regardless of their device or operating system.

Get Started Today!

Ready to experience the power of Maple AI on your desktop? Head over to our website at maple.ai/downloads to download our macOS or Linux app.

If you're new to Maple AI, sign up for a free account and start exploring the possibilities of secure and private AI. We're excited to have you join our community and look forward to continuing to innovate and improve our product to meet your needs.

What is Maple AI?

Maple AI is an AI chat tool that enables private conversations with a general-purpose AI assistant. Unlike other AI tools, Maple is built on a foundation of security and privacy, ensuring that your conversations remain confidential and protected.

Key Features of Maple AI

• End-to-end encryption: Your conversations are encrypted on your device, sent to our servers, and processed in an encrypted form by the AI.
• Secure enclaves: We use secure enclaves to manage your private key encryption, ensuring your data is protected.
• Syncing across devices: Your chats are synced automatically between devices, so you can pick up where you left off on any device, whether it is mobile or desktop.

Getting Started with Maple AI

1. Sign up for an account: Head over to maple.ai and sign up for a free account. You can choose to start with a free plan or upgrade to a paid plan for more comprehensive features.
2. Log in and start chatting: Once you've signed up, start chatting with the AI assistant. You can ask questions, provide information, or simply have a conversation.
3. Explore the interface: Familiarize yourself with the Maple AI interface, which includes features like chat history, search, and usage tracking.
4. Try out the AI: Ask the AI assistant a question or provide some information, and see how it responds. You can use Maple AI for personal or work-related tasks, such as writing, research, or brainstorming. We have lots of tips and tricks in our Maple Insights videos on YouTube.
5. Desktop Apps: If you prefer a native app approach, we have desktop apps for macOS and Linux. Download them here. (Other platforms, including Mobile, are coming soon.

Example Use Cases

Personal Use

1. Brainstorming Ideas

Maple AI is great for brainstorming ideas. Ask the AI assistant to help you develop ideas on a particular topic, and it will respond with suggestions. You can then use these ideas as a starting point for your own brainstorming session. For example, you could ask, "What are some ideas for a summer vacation?" and the AI assistant will provide you with a list of potential destinations and activities.

2. Tutoring on Difficult Coursework

Maple AI can also help with challenging coursework. Ask the AI assistant to explain a concept or topic, and it will provide a clear and concise explanation. You can ask follow-up questions or provide additional information to help the AI assistant tailor its explanation to your needs.

Professional Use

3. Writing Documentation

Maple AI can assist with writing documentation for work projects. Ask the AI assistant to help you write a document on a particular topic, and it will provide a draft you can review and edit.

Here's how it works:

1. Give Maple information about your project, this can be in the form of blog posts, code, or your own written explanation.
2. Provide relevant information or context.
3. Review and edit the draft document.

4. Creating Presentations

Maple AI can also help create presentations for work. Ask the AI assistant to help you create a presentation on a particular topic, and it will provide a suggested outline and content. You can then use this as a starting point for your own creation.

Here's how to get started:

1. Ask the AI assistant to help you create a presentation outline.
2. Provide relevant information or context.
3. When you have your outline fine-tuned, ask it to help you develop detailed content for each slide individually. This can be suggested titles, body text, and suggestions about what kinds of visuals would help explain the concept.

Want to See Maple AI in Action?

For a more in-depth look at Maple, check out our video walkthrough, where we explore the platform's features and capabilities. In this video, we'll show you how to use Maple AI for both personal and work-related tasks, and demonstrate the power of secure and private AI.

This video is part of a YouTube Series called Maple Insights, which has additional videos showing how to utilize Maple to its fullest. More videos are added regularly. Check it out here: https://www.youtube.com/playlist?list=PLgwB-Ekpi4FiNwglx5anvirwa6eJ4FwJn

Pricing Plans

Maple offers a range of pricing plans to suit different needs and budgets. Our plans include:

• Free: Access to the basic model
• Pro: $20/month
• Max: $100/month
• Team: $30/month/seat

Conclusion

Maple AI is a game-changing tool that combines the power of AI in the cloud with the privacy of AI on your device. With its end-to-end encryption, secure enclaves, and syncing across devices, Maple AI is the perfect solution for individuals and organizations looking to boost their productivity while protecting sensitive information.

Sign up for a free account today and start experiencing the benefits of confidential AI.

In only two months since its launch, Maple AI has become an essential tool for individuals from diverse professions and backgrounds.

Some of the professions represented

These are just the ones who decided to let us know what they do. We can't see anyone's data.

Our users have come to rely on Maple as a daily companion for productivity, creativity, and problem-solving. Whether dealing with privileged client information or their own, our users know Maple can guard it.

Today, we're proud to share the new Maple. We're not just building another AI; we're setting a new benchmark for the industry—one that prioritizes user privacy and security in AI innovation. Maple has big plans in its future for consumers, businesses, and app developers – a more private AI ecosystem is ahead.

What Sets Maple AI Apart

At the heart of our app lies a robust security framework built on the principles of end-to-end encryption, encrypted sync, and encrypted AI. Conversations with AI remain confidential, even from us. They are encrypted on your device using a personal encryption key unique to your user before being sent to Maple servers. At that point, only the secure enclave can interact with your data.

With a combination of our open-source server code, confidential computing secure enclaves, and reproducible builds, we provide cryptographic proof that the code running on our servers is the same as the open-source code available for review.

New Features and Enhancements

We've recently shipped exciting new features, including:

• Search your chat history: Easily find and revisit previous conversations with AI.
• AI naming of chats: Let our AI assistant suggest names for your chats, making it easier to organize and find specific conversations.
• Rename chats yourself: Take control of your chat names and customize them to fit your workflow.

Top Requests

We are listening to your feedback. Here are some of the features our users request the most. Don't take this as an announcement, but be on the lookout for some in the near future. 👀

• Apps for mobile and desktop
• Multiple LLM models
• Document upload
• System prompts
• Chat organization
• MCP interoperability

Join the Future of Private AI

We invite you to join us as we embark on this new phase. With Maple AI, you can enjoy AI's benefits without compromising your data privacy. Our commitment to transparency, security, and innovation has earned the trust of our users, and we're excited to continue evolving with you.

Sign up now and experience the power of private AI for yourself.

Pricing Plans

We offer a range of pricing plans to suit different needs and budgets:

• Free: $0 (10 chats per week)
• Starter: $5.99/month
• Pro: $20/month
• Team: $30/month/seat

Head over to maple.ai and get started today!

Developers, Get Involved!

If you're an app developer, you can build an app like Maple using our secure and private infrastructure. OpenSecret provides secure auth, private key management, encrypted data sync, private AI, and more. Join our community and start building a more secure and private future for your users.

Stay tuned for more updates, and thank you for being part of the Maple AI community!

At OpenSecret, our primary goal is to provide a secure, privacy-preserving experience for individuals and organizations using software. Our Maple app is no different. Maple is designed to help you boost productivity, whether for work or personal tasks, while ensuring that your usage of AI remains confidential and protected. With features like end-to-end encryption, automatic syncing across devices, and pricing plans to suit different needs, we're committed to making Maple AI accessible to everyone.

This blog will feature regular articles, tutorials, and success stories to help you make the most of the app. You can expect to find content on:

• Tips and best practices for maintaining online privacy and security
• Industry trends and news related to AI, chat, and cybersecurity
• Success stories and case studies of businesses and individuals who have benefited from using Maple AI
• App updates, new feature announcements, and behind-the-scenes insights into our development process
• Tutorials and how-to guides on using Maple AI's features to maximize your productivity and security

Our goal is to create a resource that not only helps you get the most out of Maple AI but also provides a platform for discussing the latest developments and challenges in AI and cybersecurity.

Whether you're a current user or just interested in learning more about using AI while handling confidential information, we invite you to subscribe to our blog and follow us on social media. We'll be posting new content regularly.

Thank you for joining us on this journey.

🤘

Get started with Maple AI today and discover a new way to work and communicate with confidence.