Pocket LLM
A real large language model, running entirely inside this browser tab on your GPU. No server, no API key, no sign-up. Your conversation never leaves your device, and once a model downloads it keeps working with your internet off.
🔒 100% private — nothing you type is ever sent anywhereHow it works
Pick a model and press Load. The first time, your browser downloads the model's weights from a public model host and caches them, which takes anywhere from a few seconds to a couple of minutes depending on the model size and your connection. After that, every reply is generated locally on your graphics card through WebGPU. There is no backend, no queue, and no per-message cost, because the inference happens on your own hardware.
How is a full LLM running in my browser?
Modern browsers expose WebGPU, a low-level API for running compute on the graphics card. The open-source WebLLM engine takes a quantized language model and runs the exact same matrix multiplications a server would, except on your GPU inside the tab. The model is loaded into a background worker so the page stays responsive while it thinks, and the weights live in the browser's cache so you only download them once. That is the whole trick: no clever server somewhere, just your own machine doing the work.
Why run a model locally?
Privacy is the obvious win. Nothing you type is transmitted, so it is safe for drafting sensitive emails, working through ideas, or pasting in text you would never send to a third-party API. It also keeps working on a plane or anywhere without signal, costs nothing per message, and has no rate limits beyond what your own hardware can do. Smaller models are fast and run on ordinary laptops; larger ones are sharper but need more memory.
Good things to try
Ask it to rewrite a paragraph in a friendlier tone, brainstorm names for a project, explain a concept simply, draft a reply, summarize notes you paste in, or write a small code snippet. The persona buttons set a system prompt for common jobs, and you can write your own. Share your exact setup with a teammate using the persona link, which carries your chosen model and system prompt but never your conversation.
Does my conversation get sent to a server?
No. The model runs in your browser on your GPU. Your prompts and the replies never leave your device. The only network request is the one-time download of the model weights.
What do I need to run it?
A browser with WebGPU, such as a recent desktop Chrome or Edge, and some free memory. Smaller models need a few hundred megabytes; larger ones a couple of gigabytes. If your browser does not support WebGPU, the page will tell you.
Does it really work offline?
Yes. Once a model has downloaded, its weights are cached. You can go offline, reload, and keep chatting. Cached models are marked in the picker.
Why is the first load slow?
It is downloading the model, which can be hundreds of megabytes. That happens once per model, then it is instant on future visits because the weights are cached locally.
Is it as good as ChatGPT?
These are small models chosen to run on consumer hardware, so they are not as capable as the largest hosted models. They are genuinely useful for rewriting, drafting, brainstorming, and quick questions, with the trade-off being full privacy and no cost.
Is it free?
Completely free, no account and no API key. Because it runs on your own machine, there is nothing to bill.
I build tools like this every day.
Senior full-stack engineer, available for senior or contract work, fully remote. See the rest of the lab or get in touch.