Why On-Device Function Calling Matters

The real magic of AI happens when a model stops merely describing the world and starts interacting with it. With on-device function calling, your app can parse natural language commands and invoke OS-level actions — instantly, privately, and without internet dependency. Google's AI Edge Gallery now makes this accessible on both Android and iOS.

Key capabilities unveiled:

  • Mobile Actions demo: Voice assistant that creates calendar entries, navigates maps, or toggles flashlight — all offline.
  • Tiny Garden demo: Interactive game where voice commands like "Plant sunflowers in the top row" are decomposed into custom app functions.
  • Cross-platform launch: iOS app now available on App Store, bringing the same agentic features to Apple hardware.

For a deeper look at semantic UI patterns, check out our guide on Building the Perfect Pie Chart in CSS.

Smartphone running Google AI Edge Gallery with on-device function calling demo Coding Session Visual

Performance Benchmarks & How to Run Your Own

Using Mobile Actions as an example, the performance is blazingly fast on CPU — clocking in at 1916 tokens/sec (prefill) and 142 tokens/sec (decode) on a Pixel 7 Pro. Here's how to benchmark on your own devices:

# Clone the Gallery app repository
git clone https://github.com/google-ai-edge/gallery-app
cd gallery-app

# Build and install on your Android device
./gradlew installDebug

# Run the benchmark from the app menu
# Results will be displayed in the console

Note: iOS benchmarking is coming soon. Stay tuned for updates.

Developer using AI agent on mobile for natural language command parsing Dev Environment Setup

Limitations & Considerations

While on-device function calling is revolutionary, keep these caveats in mind:

  • Model size: The 270M FunctionGemma is compact, but still requires ~150MB storage.
  • Battery impact: Continuous voice processing can drain battery faster than traditional app interactions.
  • Accuracy: Complex multi-step commands may still fail; always provide fallback UI.
  • Privacy: All processing stays on-device, but model updates require periodic downloads.

For a related exploration of domain-specific AI in healthcare, read Bridging AI and Medicine with Claude in Microsoft Foundry.

Cross-platform development with Google AI Edge on laptop and phone System Abstract Visual

Next Steps: Build Your First Local Agent

Ready to dive in? Follow these steps:

  1. Download the AI Edge Gallery from Google Play or App Store.
  2. Explore the demos — Mobile Actions and Tiny Garden are great starting points.
  3. Fine-tune FunctionGemma for your own app logic using the provided SDK.
  4. Deploy on both platforms using the cross-platform Google AI Edge stack.

What to learn next:

  • Study the FunctionGemma model architecture to understand token prediction.
  • Experiment with custom function definitions in your app's intent system.
  • Join the Google AI Edge community for early access to iOS benchmarking.

We can't wait to see the agentic features you'll bring to life. Happy coding!

This content was drafted using AI tools based on reliable sources, and has been reviewed by our editorial team before publication. It is not intended to replace professional advice.