How a private AI chat works
The page downloads a small language model once and runs it in your browser, on your graphics card where the browser allows it, otherwise on your processor. Type a message and the answer appears word by word.
Pick a model
Better is Qwen2.5 0.5B: clearer answers, and it can write in other languages. Fast is SmolLM2 360M: a smaller download and quicker answers, in English only.
Send a message
The first time, the model downloads (290 to 800 MB) and is kept on your device. After that it starts in a few seconds, even without internet.
Keep talking
It remembers the conversation until you start a new chat or close the page. Very long chats drop their oldest messages to make room.
What it is good at, and what it is not
These are small models: about 200 to 400 times smaller than the large online assistants. They are good at explaining everyday things, drafting short messages, brainstorming names and ideas, and rephrasing. They are not reliable for facts, numbers, recent events, maths or medical, legal and money questions: they can sound sure and still be wrong.
For summaries and rewriting there are tools made for it: Summarise text and Rewrite text use the same models with instructions tuned for the job.
Where do my messages go?
They stay on your device. The AI model downloads from this site once and runs right in your browser. Your chat is gone when you close the tab.
- Downloaded once, then kept in your browser for next time
- No account and no message limit
Models and licences
| Better | Qwen2.5 0.5B Instruct by the Qwen team at Alibaba Cloud, Apache-2.0. 786 MB on a graphics card (4-bit), 512 MB on a processor (8-bit), in onnx-community's ONNX export |
|---|---|
| Fast | SmolLM2 360M Instruct by Hugging Face, Apache-2.0. 273 MB on a graphics card (4-bit), 365 MB on a processor (8-bit) |
| Runtime | onnxruntime-web (MIT), 14 or 17 MB |
Specs and browser support
| Speed | Roughly 10 to 40 words a second on a processor, faster on a recent graphics card |
|---|---|
| Memory | About 4,000 tokens (3,000 words) of conversation; older messages are dropped past that |
| Browsers | Current Chrome, Edge, Firefox and Safari on a computer. Graphics card use needs WebGPU with 16-bit floats (Chrome and Edge on most computers) |
| Price | Free, no account, no limits |
Questions
No. The AI model is downloaded to your browser and runs there. What you type and what it answers stay on your device, and the conversation is gone when you close the page.
Much weaker. These models are 200 to 400 times smaller than the big online assistants, so they can be wrong about facts and lose track of long questions. They are good at short answers, explanations, drafting and brainstorming.
The first time, your browser downloads the model: about 290 to 800 MB on a graphics card, or 380 to 530 MB on a processor. It is kept on your device, so the next visit starts in seconds.