The true potential of AI unfolds not when it describes the world, but when it acts upon it. A key mechanism for this action is function calling—the ability for a model to interpret natural language and invoke specific tools, like opening an app or adjusting a setting. Traditionally, this required hefty models and a server connection. Google AI Edge Gallery's latest update changes the game by bringing sophisticated, low-latency function calling entirely on-device. Let's dive into what this means for the future of mobile interaction. Reference Material
Mobile Actions: Redefining the Offline Assistant
The newly introduced Mobile Actions demo is powered by a compact model called FunctionGemma. It runs locally on your phone, parsing commands like "Show me the San Francisco airport on map" or "Turn on the flashlight," and maps them to the correct OS-level action—all without an internet connection.
Why it matters:
- Instant Response: No network round-trip latency.
- Enhanced Privacy: Your data never leaves the device.
- Always Available: Works in areas with poor or no connectivity.
This shift enables a new class of responsive and private voice assistants and agentic features directly in your pocket.
Feature Breakdown & Cross-Platform Reach
| Demo / Feature | Core Model | Key Capability | Use Case Example |
|---|---|---|---|
| Mobile Actions | FunctionGemma | Natural Language → System Function (Offline) | Create calendar events, navigation, device control |
| Tiny Garden | FunctionGemma (270M) | Natural Language → Custom App Logic | "Plant sunflowers in the top row and water them" |
| AI Chat / Ask Image | Various On-Device Models | Multi-turn chat, image query | Local, private conversations and image analysis |
Now on iOS
A major milestone of this update is the launch of the Google AI Edge Gallery app on the iOS App Store. iOS developers and enthusiasts can now experience the same rich on-device features, including the agentic Mobile Actions and Tiny Garden demos, showcasing seamless tool-calling performance on Apple hardware.
Getting Started for Developers
To adapt this for your own use cases, you can fine-tune your own version of FunctionGemma using the Google AI Edge stack. The Gallery app also includes benchmarking tools, allowing you to test model performance (e.g., tokens/sec) directly on your hardware. For instance, Mobile Actions runs at a blazing 1916 tokens/sec (prefill) on a Pixel 7 Pro's CPU.
Key Takeaways:
- On-Device is Viable: Complex function calling is no longer server-bound.
- Privacy by Design: Sensitive commands are processed locally.
- Universal Access: The technology is now accessible across Android and iOS ecosystems.
The move towards powerful on-device agents marks a significant step in making AI interactions faster, more reliable, and more private. It's an open invitation to reimagine what's possible in your mobile applications.