Gemma 4 Runs Smoothly on Android

The interesting part here isn’t the model size, it’s the runtime stack. On mobile, the difference between “technically works” and “actually usable” is often whether you can hit GPU- or NPU-aware inference paths instead of burning the CPU.

• –llama.cpp confirms the usual mobile ceiling: portable, familiar, but too slow for real interactive use on this device
• –LiteRT-LM changes the equation by using Android-optimized execution, which is what makes the same model feel smooth
• –Wrapping inference behind a local HTTP server is the right integration move because it turns a phone model into a tool-callable backend
• –This is a strong pattern for private, offline, on-device agents where latency and data locality matter more than raw benchmark scores
• –The writeup is more useful as an Android deployment playbook than as a Gemma benchmark