What's this about?
Cactus taught Gemma 4 (a small on-device model) to attach a confidence score to every answer, so a phone can answer locally when confident and hand off to a big cloud model only when unsure.
What exactly is "on-device AI"?
Running the model on your phone/laptop itself instead of a datacenter — private, free, and offline, but smaller and dumber. The hard problem is knowing *which* questions the small model will fumble. Calibrated confidence is the missing piece that makes local-first, cloud-fallback practical.
Why is it worth noting?
It's a clean, open implementation of the hybrid pattern everyone (Apple included) is converging on, with copy-paste quickstarts for llama.cpp, MLX, and Transformers.