AI in your browserHow to run an AI model in your own browser

Open-weight language models are now small enough to run inside a web page, on your own computer's graphics hardware. Nothing you type is sent anywhere. What that takes, what it costs, and how to check it for yourself.

A guide from Zero Witness · September 27, 2026

The short answer

Use a page that runs an open-weight model in the browser with WebGPU. The model is downloaded once (from a few hundred megabytes to a few gigabytes), kept in your browser’s storage, and run on your own graphics hardware inside the tab. Your messages are never sent. On Zero Witness, choose “In your browser” and pick a model. It’s free and needs no account.

What is actually happening

A language model is a very large file of numbers and a short program that does arithmetic with them. “Open-weight” means its maker has published those numbers for anyone to run. Compressed (quantised) versions of the smaller ones fit in a few hundred megabytes to a few gigabytes.

WebGPU is the part of a modern browser that lets a web page use your graphics hardware. With it, a page can do that arithmetic on your own machine. The model goes into the storage your browser keeps for that one site. Nothing is installed: no program, no icon, and clearing the site’s data removes it.

What your device needs

  • A browser with WebGPU and half-precision support. The models here need both. If yours lacks either, the page says which, and offers our hardware instead.
  • Room in the browser’s storage for the model you choose. The page checks before it downloads anything and names the figures if there isn’t.
  • On a phone or tablet, only the smaller models. The page offers only those that fit.

Speed can’t be predicted in advance. Browsers deliberately don’t tell a web page how much graphics memory a device has, so a page that promises a speed before trying is guessing. What can be checked is checked; the rest you find out on the first answer.

Which model to start with

Zero Witness offers ten, from Alibaba (Qwen), Meta (Llama), Google (Gemma), Microsoft (Phi) and Hugging Face (SmolLM), from about 200 MB to about 2.6 GB. Qwen3 1.7B, about 1 GB, is the one it starts with. Qwen2.5 Coder is aimed at code, and Phi-3.5 Vision is the only one that reads images.

Bigger is not always better. In our own tests, one of the larger models stated made-up recent news as fact where a smaller one said plainly that it could not know. The benchmarks page has what we measured, including the failures.

What a small model is good and bad at

  • It gets more wrong than the large models behind commercial assistants. It’s fine for drafting, rewording, explaining and brainstorming; check anything factual.
  • It holds a limited amount at once, roughly 1,700 words of conversation and document. Past that, the earliest turns fall out of what it can see, and the page says so when it happens.
  • It has no internet. It can’t know today’s news unless you turn on web lookups, and those send the search terms (never the conversation) to a search engine.

How to check that nothing is sent

  • Watch the network. Open your browser’s developer tools (F12, or Cmd-Option-I on a Mac), choose the Network tab and send a message. The first time you’ll see the model’s files load from our site. After that, no request carries what you typed.
  • Turn off your Wi-Fi. Once the model has answered one message, turn off your Wi-Fi and, without reloading the page, send another. It still answers. Reloading the page needs the connection again, because the page itself comes from the web.

The network tab will also show small page-view signals sent to our own site. They say that somebody is using the page. They carry nothing you typed.

How to start on Zero Witness

  1. Open the chat.
  2. Press the pill beside Send that says where the model runs, and choose “In your browser”.
  3. Pick a model, or keep the one suggested for your device.
  4. Send a message. The first time, the model loads, with its progress shown. After that it starts at once on this browser.

One tip for a large model: bookmark the site first. Some browsers only agree to keep a site’s storage for good once you have bookmarked it or used it often, and without that they may stop a large download partway.

See how these models perform. Related guides: How to tell whether an AI chat is really private · How to transcribe a recording without uploading it · How to turn a photo into a PDF without uploading it. All guides.

Questions

In your browser, yes: the model runs on your device and no request carries what you type. You can check that yourself with the network tab or by turning off your Wi-Fi. If you choose our hardware instead, your messages are encrypted in your browser and read in memory on hardware we own, and that is a promise about how we run it rather than something you can check.

No. There’s no account and no message limit in your browser, because your own device does the work.

It runs a much larger model than fits in a browser tab, holds more conversation at once, and has nothing to download. The cost is that your words leave your device, encrypted.

Yes. Settings, then Privacy, lists every model this browser holds with its size and a button to remove it. Clearing the site’s data in your browser removes them too.

The smaller models can, on a recent phone with a browser that supports WebGPU. The page checks and offers only the ones that fit.

Open the chat Free. No account.