unsloth/functiongemma-270m-it

Overview

FunctionGemma-270m-it is a compact, instruction-tuned model from Google DeepMind, specifically engineered for function calling. Built upon the Gemma 3 270M architecture, it leverages the same research and technology as the Gemini models. Unlike general dialogue models, FunctionGemma is intended as a foundation for fine-tuning to create highly specialized function-calling agents, particularly for text-only interactions.

Key Capabilities

• Function Calling Specialization: Designed from the ground up for function calling, translating natural language into structured function calls.
• Resource Efficiency: Its small 270 million parameter size and 32K token context length make it suitable for deployment in resource-constrained environments like laptops, desktops, or on-device applications.
• Fine-tuning Potential: Highly performant after further fine-tuning for specific function-calling tasks, including multi-turn scenarios.
• Versatile Deployment: Optimized for various hardware, demonstrating strong performance in single-turn scenarios, with enhanced accuracy post-fine-tuning on task-specific data.

Use Cases & Performance

FunctionGemma has been showcased in two primary use cases:

• Tiny Garden: A fine-tuned model powering a voice-controlled interactive game, managing game logic by decomposing commands into app-specific functions.
• Mobile Actions: Translates user inputs into Android OS system function calls, enabling offline, private agent capabilities for personal device tasks. Fine-tuning significantly boosts performance, with a base model achieving 58% accuracy on Mobile Actions, improving to 85% after fine-tuning.

Training involved 6T tokens, including public tool definitions and tool use interactions, with a knowledge cutoff of August 2024. On-device performance evaluations on a Samsung S25 Ultra show efficient prefill and decode speeds with low memory footprints for fine-tuned applications.