On-Device Function Calling in Google AI Edge Gallery

This post announces major updates to Google AI Edge Gallery, showcasing on-device function calling powered by the compact FunctionGemma model on mobile hardware.

• •FunctionGemma, a 270M parameter model, translates natural language into function calls entirely on-device without any server connectivity
• •Mobile Actions demo parses commands like "Show me SF airport on map" or "Turn on the flashlight" into OS tool or app intent invocations offline
• •Tiny Garden demo shows the model handling custom app-specific function calls (plantCrop, waterCrop with grid coordinates) for a voice-controlled mini-game
• •Google AI Edge Gallery is now available on iOS in addition to Android, with multi-turn AI Chat, Ask Image, and Audio Scribe features included
• •Built-in benchmarking is now integrated into the app; FunctionGemma achieves 1916 tokens/sec (prefill) and 142 tokens/sec (decode) on a Pixel 7 Pro