FunctionGemma

An open model specialized for function calling at the edge

Develop fast, private, local AI agents with a specialized version of Gemma 3 270M

Unified action and chat

Generates function calls to execute tools, then switches context to summarize the results in natural language

Efficient multi-model orchestration

Process common commands on-device or route to larger models for more complex tasks

Optimized for local deployment

Enables inference on mobile and IoT devices so data is processed locally, protecting user privacy

Built for customization

FunctionGemma is designed as a base for further training and can provide an efficient pathway to production-grade performance.

In our evaluation on the "Mobile Actions" dataset, fine-tuning improved accuracy by 47% over the base model. Try it in Colab.

A bar chart showing FunctionGemma accuracy on Mobile Actions tasks, comparing the base FunctionGemma 270M model at 58% accuracy to the fine-tuned FunctionGemma 270M model at 85% accuracy.A bar chart showing FunctionGemma accuracy on Mobile Actions tasks before and after fine-tuning. The Base FunctionGemma 270M model achieves 58% accuracy, while the Fine-tuned FunctionGemma 270M model achieves 85% accuracy.

Demos