Gemma 4 runs well on Android

This is a runtime story, not a model story. The interesting part is that LiteRT-LM turns “technically works” into “actually buildable,” which is what edge AI on Android has been missing.

• –llama.cpp on Termux is the baseline many people try, but the CPU-only path makes Gemma 4 feel unusably slow on phones
• –LiteRT-LM is the unlock here because it can use Android hardware more intelligently, including GPU delegation where it helps
• –Bridging the local model into Termux gives you a useful architecture: on-device inference plus scriptable agent control
• –ADB automation makes this more than a chat demo; it becomes a phone that can act on its own apps while staying offline
• –The broader implication is that “local assistant on mobile” is becoming a systems-integration problem, not a model-size problem