This is the actual LumaBrowser chat interface, not a video and not a mockup. Send a message and a small language model loads once, then runs entirely in your browser tab on your own GPU through WebGPU. There is no server doing the thinking, so nothing you type ever leaves your machine.
Nothing you type leaves your device. The model runs in your browser tab, so your messages are never sent to an inference server. The only network traffic is the one-time model download, which your browser then caches, so a second visit starts without downloading anything again.
There is no account and no setup. The first message does the one-time work; everything after that is instant.
Launch it and you land in the same chat UI the desktop app uses, with a few example conversations and prompt suggestions to get going.
Your first message triggers a one-time model download of about 450 MB. A progress bar shows it loading into your browser.
The model runs locally through WebGPU. Replies stream back token by token, generated on your hardware, not ours.
Your browser caches the weights, so the next visit skips the download and starts chatting right away.
The demo runs a compact model so it fits in a browser tab. It is chat-only by design. The full desktop app is where LumaBrowser opens up.
Because the inference happens on each visitor's own GPU, there is no per-message or per-visitor server bill behind this page.
The chat surface here is the same one shipped in the desktop app, including streaming replies and rich message rendering.
No sign-up, no server-side history, no prompt logging. The conversation lives in your browser and nowhere else.
The browser demo is deliberately minimal. The installed app runs much larger local models and adds the capabilities LumaBrowser is built around.
| This browser demo | Desktop LumaBrowser | |
|---|---|---|
| Model size | One small model (about 0.6B) that fits a browser tab | Curated multi-billion-parameter local models sized to your hardware |
| Runtime | WebGPU inside your browser | Managed llama.cpp server on your GPU, with CUDA, Vulkan, and Apple Silicon |
| What it does | Chat only | Chat plus browser automation tools, sub-agents, knowledge bases, and local image generation |
| Setup | None, it runs in the page | A one-click guided setup that picks a model to fit your machine |
Want the bigger models and the full toolset? See Local AI, AI Sub-Agents, Knowledge Base, and Local Image Generation.
The demo relies on WebGPU to run the model. If your browser does not expose it, the demo shows a short notice instead of failing silently.
Recent desktop Chrome and Edge support WebGPU out of the box. This is the smoothest way to try it.
Newer versions are adding WebGPU support. If yours has it, the demo will run.
The model runs on your graphics hardware, so a desktop or laptop with a modern GPU gives the best experience.
You will see a friendly notice with a link to the desktop app, which runs locally without any browser requirement.
Try the chat in your browser, then install LumaBrowser to run far larger local models with tools, agents, knowledge bases, and image generation. It is free to download.