Tools other people built
Anything we can run ourselves, we do, in the studios. What is left is the handful we genuinely cannot host yet: a different inference engine, or kernels nobody has published a licence for. These open on their own sites.
Tools built by other people
- Bonsai Image: Text to image at 1024px, squeezed into 2-bit weights with hand-written WebGPU kernels. A bet that crushing a 4B model down to two bits beats running a smaller one.
- WebLLM Chat: A full local chat workspace on a different engine to ours, with a very wide model picker.
- Gemma 4 WebGPU: Google's open Gemma 4 handling text, images and audio at once. We may bring this in-house later.