Run Llama 3 in your browser

Meta's Llama 3.2 1B is one of the models you can pick in the chat on this site. It downloads once, about 880 MB, and then answers entirely on your own machine — no server, no API key, no account.

Try it in the chat Questions and troubleshooting Why the Neural Engine isn't used

What your device needs

Llama 3.2 1B runs on your graphics chip through WebGPU, the browser standard that gives a web page access to graphics hardware. That support is what decides whether this model is available to you.

Chrome and Edge: supported

Safari 18 and later: supported

Older browsers: processor-only mode instead

Roughly 2 GB of free memory keeps it comfortable. If your browser has no graphics acceleration, the chat quietly falls back to a much smaller model that runs on the processor — slower, but it still works.

What to expect

DownloadAbout 880 MB, once. Kept in the browser afterwards.

SpeedA few words per second on a tablet; considerably quicker on a desktop.

QualityNoticeably better conversation than the smallest models, still a 1-billion-parameter model — keep questions short and concrete.

PrivacyNothing you type leaves the device.

OfflineOnce downloaded it answers with no connection at all.

How to try it

Open the chat and pass the short human check.
Choose "Llama 3.2 1B Instruct" in the model picker.
Wait for the one-time download — on a fast connection this is a couple of minutes.
Ask it something. Everything after that happens on your device.

On a mobile connection it is worth waiting for Wi-Fi: the download is the same size either way.

Other models on the same page

The picker also offers Qwen2.5 0.5B for the fastest start, Qwen2.5 1.5B as a stronger all-rounder, and Phi-4 mini for the most capable answers on a recent desktop.