On this device
Models that run on your own device
Small open models download once into this browser and answer on your GPU. No key, no sign in, and nothing you type leaves the device. They cost nothing because your device does the work.
- No key, no sign in
- Your words stay on this device
- Free, because your GPU does the work
This device
Asking this browser for its GPU
6 models, small to large
Each is a download you choose, once. It stays in this browser until you delete it.
SmolLM2 360M
by Hugging Face
212 MB download · 4k context
needs about 376 MB of GPU memory
Qwen3.5 0.8B
by Qwen
443 MB download · 4k context
needs about 1.6 GB of GPU memory
a thinking model, asked to answer directly
Gemma 3 1B
by Google
602 MB download · 4k context
needs about 711 MB of GPU memory
Llama 3.2 1B
by Meta
710 MB download · 4k context
needs about 879 MB of GPU memory
Qwen3 1.7B
by Qwen
985 MB download · 4k context
needs about 2 GB of GPU memory
a thinking model, asked to answer directly
Llama 3.2 3B
by Meta
1.8 GB download · 4k context
needs about 2.3 GB of GPU memory
How it works
The engine is WebLLM by MLC (Apache 2.0), version 0.2.85. Each model is compiled ahead of time into WebGPU programs and runs in a worker beside this page.
The weights come from Hugging Face and the compiled model from GitHub, once. This browser keeps them in its Cache Storage until you delete them here or clear the site's data.
Your GPU writes every reply. No prompt and no answer goes to Antimatter or to any model provider. In chat, a thread is saved the way your other threads are.
Bigger models live in the cloud, behind one key. Open chat