Open-weight intelligence that runs entirely on your hardware. No cloud calls, no API keys, no data leaving the room.

Each one talks to your local Gemma 4 through a same-origin proxy. Play them right here.
ollama pull gemma4:e2b-it-qat — update Ollama, then pull the model.
QAT — quantisation-aware training — teaches Gemma 4 to stay sharp at lower precision. The result on a real M2 MacBook Air: smaller on disk, faster to respond, and quality that matched or beat the non-QAT build.
Three steps from zero to a private model humming on your own machine.
Grab the latest build so it knows the Gemma 4 tags. One download, runs in the background.
Pull the QAT weights — 4.3 GB on disk, tuned for everyday laptops. Then it's yours, offline.
Refresh this page. The status light turns cyan and the demos above go live, instantly.
ollama pull gemma4:e2b-it-qat