Local AI chat

model not loaded

A private chat in your browser

Pick a model at the bottom right — it downloads once and stays in your browser.

Ask anything — the answer is generated right on your device.

WebGPU required

Your browser does not support WebGPU. Open this page in a recent Chrome, Edge, or another browser with WebGPU enabled.

Prompt catalog · All tools

The AI model can make mistakes — verify important information.

How to run a local AI model in your browser

Local AI chat is a language model that runs right on your computer, in the browser, via WebGPU. Unlike cloud chatbots, your messages aren't sent to an AI server: your device's GPU generates the response. You can choose from open models: Qwen, Llama, Gemma, Phi and DeepSeek R1.

How to use it

Open the page in an up-to-date Chrome or Edge on a computer. WebGPU support is required; if it's missing, the tool will tell you.
Click "Model" in the bottom right and pick one. The recommended Qwen2.5 1.5B is a good place to start.
Wait for it to load. Models range from ~0.9 to ~5 GB, download once, and stay in your browser's storage.
Type your question and send it. You can stop a response with the same button.
To start a new conversation, click "New chat". Past chats stay in the "My chats" list on the left: you can open and continue them, rename or delete them.

Which model to choose

Low-end laptop or integrated graphics — Qwen2.5 0.5B or Llama 3.2 1B (~0.9 GB).
Typical laptop with 8 GB of RAM — Qwen2.5 1.5B, Gemma 2 2B or Qwen3 1.7B (1.6–2 GB).
16 GB of RAM or more, or a gaming GPU — Llama 3.2 3B, Qwen2.5 3B, Qwen3 4B, Phi 3.5 mini (2.3–3.7 GB).
Powerful GPU or a Mac with 32 GB — Qwen2.5 7B or DeepSeek R1 8B (~5 GB). Qwen3 and DeepSeek R1 show their reasoning first.

Settings and tips

Temperature from 0 to 1.5: lower is more precise and predictable, higher is looser and more creative. The default is 0.7.
You can attach up to 5 text files, up to 256 KB each: TXT, Markdown, CSV, JSON, code. The model reads them as part of your question.
If responses are slow, pick a smaller model. If they're fast, try a bigger one.

New and saved chats

New chat — the button at the top left. It starts a clean conversation without losing the previous one.
My chats — the list under the button. Every conversation is saved automatically after each message; its title comes from the first question.
Click a chat in the list to open and continue it: the model gets the whole conversation again and "remembers" what it was about.
You can rename a chat (pencil) or delete it (bin). Next time you come back, the chat you were in last opens.
Where chats are stored: only in your browser on this device (IndexedDB storage). They are never sent to our server. If you clear the site data in your browser or open mozgi.io in another browser or on another device, these chats won't be there.

Limitations

These are small models: they're noticeably weaker than large cloud chatbots, make mistakes, and don't know recent news. Double-check anything important.
The model can't access the internet and doesn't understand images.
Chats are saved only in this browser on your device. If you clear the site data or open the page in another browser, they won't be there.
You need an internet connection to open the page and download the model. Once the model is loaded, responses are generated without going online.
Not all phones support WebGPU, and they often don't have enough memory.

Privacy

Your messages and responses are processed on your device and aren't sent to our server. Only the model weights (from Hugging Face) and the WebLLM library (from jsDelivr) are downloaded from the internet. The site collects visit statistics: Cloudflare Web Analytics without cookies, while Yandex Metrica and Google Analytics load only if you've accepted cookies. Conversation text is excluded from analytics recording. Learn more in our privacy policy.

More on mozgi.io: the video prompt generator, the AI prompt catalog, and skills for Claude and AI agents.

Local AI FAQ

Is this like ChatGPT, but on my own computer?

In principle, yes: it's a chat with a language model. But the models here are small, so their answers are simpler and they make mistakes more often. On the upside, your messages aren't sent to an AI server.

Do I have to download the model every time?

No. The model stays in your browser, and next time you just select it — it loads from the cache in a few seconds. If you clear the site data, you'll need to download it again.

Why does it say "WebGPU required"?

Your browser or GPU doesn't support WebGPU. Update Chrome or Edge on your computer.

Are conversations saved?

Yes, but only in your browser. Each chat is saved automatically and stays in the "My chats" list on the left: you can open and continue it, rename or delete it. Conversations never go to a server, so they won't be on another device or after you clear the site data.

Do I need a subscription?

For the first hour you can use it without signing up or subscribing — the hour counts from the first time you open any of the locked mozgi.io tools. After that you need a subscription: it gives full access to all locked tools, prompts and skills. Your saved chats stay in your browser either way.