Gemma 4 brings the next gen of on-device AI to Android. Get code-assistance in Android Studio and build intelligent experiences that run locally using the ML Kit GenAI Prompt API. Start building agentic experiences on-device → https://goo.gle/4sOiBAv
Tested the 27b model against Qwen 3.5 122B for Android coding in LMStudio, and it’s a lot closer than i could imagine. Incredible performance for its size—nice work!
Native tool use on 2B/4B models is a massive step for local-first agents. The industry's next challenge is the security boundary—specifically, how we audit these autonomous host-level interactions without adding latency that breaks the reasoning loop. I'm curious if the team has explored how these smaller models handle "denied" tool-call feedback from an external policy engine. Ensuring safety at the edge is going to be a collective effort. Great release.