FunctionGemma: Google’s Compact AI for Edge Workloads

What is FunctionGemma?

FunctionGemma is a 270M parameter transformer that specializes in function calling, designed to operate as an edge agent translating natural language into API actions.

Architecture and Training Data

Maintaining the Gemma 3 architecture, FunctionGemma boasts a 270M parameter scale, utilizing the same research infrastructure as Gemini. It is optimized for JSON structures with a 256K vocabulary, enhancing token efficiency for function schemas.

Training Dataset:

• 6 trillion tokens, with a cutoff in August 2024.
• Focus on public API definitions and interaction data for function calls.

Conversation Format and Control Tokens

FunctionGemma requires a strict conversation template. Roles like developer, user, or model are marked with tags such as <start_of_turn>role ... <end_of_turn>. Control tokens allow clear differentiation between natural language and function declarations, calls, and responses.

Fine Tuning and Mobile Actions Performance

While FunctionGemma delivers decent performance out of the box, task-specific fine-tuning is essential for achieving reliability in production. Accuracy can rise from 58% to 85% following fine-tuning on suitable datasets.

Edge Agents and Reference Demos

Targeted at edge agents for local execution, FunctionGemma runs effectively on devices like laptops and smartphones. Google has introduced several demos, such as Mobile Actions, which exhibit offline capabilities, and other interactive applications like Tiny Garden and the Physics Playground.

Key Takeaways

1. FunctionGemma is tailored for function calling, not general chat.
2. It operates within a 32K token context and utilizes a dedicated vocabulary.
3. The model employs a rigorous chat format for reliable tool interaction.
4. Fine-tuning enhances accuracy significantly for specific tasks.
5. Low resource requirements make it suitable for various consumer devices.