Google has released FunctionGemma, a specialized Gemma 3 270M model fine-tuned for function calling—turning natural-language commands into structured API/tool calls for local agents.
Why this matters for edge AI engineers
- Local-first “action agents”: FunctionGemma is positioned as an on-device “action” model (or a lightweight controller that can route complex requests to larger models like Gemma 3 27B).
- Built to be fine-tuned for determinism: In Google’s “Mobile Actions” evaluation, fine-tuning improved accuracy from 58% to 85%, underscoring their message that edge agents often need task-specific training for reliability.
- Designed for constrained devices + structured I/O: Google calls out edge suitability (including examples like Jetson Nano and phones) and notes the model’s 256k vocabulary helps tokenize JSON efficiently—useful when your agent is constantly emitting structured function-call payloads.
How to try it
FunctionGemma is available with open weights (licensed for responsible commercial use) and can be downloaded via common channels (e.g., Hugging Face / Kaggle). Google also points developers to a fine-tuning “cookbook” and demos via the Google AI Edge Gallery app (including “Mobile Actions” and other local demos).
Note: official docs emphasize that FunctionGemma is intended as a fine-tuneable tool-calling base (not a general chat model) with a 32K context window, and provide specific formatting tokens/templates for tool use.
Further Reading:
https://blog.google/technology/developers/functiongemma/
https://ai.google.dev/gemma/docs/functiongemma
