WebGPU Chat downloads an open language model once and runs it on your own graphics card. There is no account, no server reading what you type, and no analytics.
The website only serves static files. All the computing happens on your computer, and the browser prevents the page from connecting to any site other than itself and Hugging Face, where the models are downloaded from.
Messages, replies and history stay in your browser's storage. You can delete them whenever you want.
After the first download, the model is kept in the browser cache and you can chat without internet.
The code is on GitHub under the MIT license, and the models are open. You can check what it does and host it yourself.
The app checks your GPU, recommends the model that fits best and blocks the ones that won't fit. They are downloaded directly from Hugging Face, in a fixed and verified version.
Free use and redistribution; commercial use is only allowed for entities with less than 10 million US dollars in annual revenue. Full text.
Apache 2.0: free use, including commercial use, keeping the license notice.
Apache 2.0: free use, including commercial use, keeping the license notice.
Each model went through its own battery of 30 questions in Spanish and English (conversation, code and reasoning). Full answers and notes are in the repository.
Measurements in progress.
No inference server: the website is a set of static files, and your browser does the rest.
Reads your card's WebGPU capabilities (16-bit numbers, maximum buffer size) and recommends a model.
The weights, quantized to 4 bits in ONNX format, are fetched from Hugging Face at a pinned revision and saved in the browser cache.
Transformers.js and ONNX Runtime Web run the model with WebGPU in a separate thread, using each model's official chat template.
The key-value cache is kept between messages, so each turn only processes what is new and replies sooner.
They fit on a personal computer in exchange for knowing less than the large commercial models. They can make up data, names or figures with complete confidence. Check what matters and do not use them for medical, legal or financial decisions.
Your messages do not leave the browser. The only connection to the internet is the download of the models from Hugging Face, which, like any server, sees your IP address. The website has no analytics or cookies, and its content security policy (CSP) only allows it to connect to itself and to Hugging Face. You can check this in the "Network" tab of the developer tools.
A browser with WebGPU and a GPU that supports 16-bit numbers (shader-f16): recent Chrome or Edge on Windows, macOS or ChromeOS, Chrome on Android, or Safari 26. The light model works with integrated graphics; the maximum one needs a powerful dedicated GPU.
Between 1.1 and about 8 GB per model, in the browser storage. In the app, "Models" shows how much each one takes and lets you delete it. They also disappear when you clear the site data in your browser.
On the processor these models would produce only a few words per second, too slow for chatting. That is why the website requires WebGPU and explains what to do if your browser does not have it.
Yes. It is a personal, non-profit project, with no ads and no sign-up. The computing cost is paid by your own computer.
WebGPU Chat