Introducing FunctionGemma: A Lightweight Model for Function Calling

FunctionGemma: Specialized for Function Calling

• 💡 FunctionGemma is a specialized version of the Gemma 3 270M model, fine-tuned for function calling and tool use.
• 🎯 It's designed for developers to build fast, private, and cost-effective applications that translate natural language into API actions.
• 🚀 Despite its small size (270 million parameters), it's trained to determine the right function and can be fine-tuned for specific tasks, achieving high success rates.

On-Device Efficiency and Performance

• 📱 FunctionGemma's lightweight footprint allows it to run efficiently on embedded and mobile hardware, even on-device.
• ⚡ With accelerators like GPUs and NPUs, its processing speed for input and actions can be significantly increased.
• 🌐 This enables mobile developers to move beyond chat-based interactions and execute actions directly on the device.

Demos Showcasing FunctionGemma in Action

• 🛠️ Mobile Actions demo allows users to trigger device actions (like flashlight or calendar) via voice or text input by fine-tuning FunctionGemma on a set of tools.
• 📈 Using the new function calling format improved accuracy over the base Gemma 3 27 model, with further improvements after fine-tuning.
• 🎮 An interactive farming game demo showcases how a fine-tuned FunctionGemma model can drive game mechanics from natural language commands, like planting and watering crops.

Use Cases and Developer Resources

• 🧩 Functionality to generate function calls and arguments from free-form input enables use cases like data lookup, routing queries, and new interaction modalities for games and apps.
• 🔒 Running AI on-device with FunctionGemma unlocks benefits such as privacy, offline capabilities, and reduced cloud costs.
• 📚 Developers can fine-tune their own specialized function calling models and find resources like Colab notebooks and the Google AI Edge Gallery app to get started.
• ☁️ FunctionGemma is available on platforms like Hugging Face, Kaggle, and Vertex AI, and is compatible with popular tools and frameworks.