Function Gemma

WIP nlp-llms function-calling gemma language-models llm nlp on-device 1 min read

Google's open-weights LLM specifically tuned for function calling and on-device agents

Overview

FunctionGemma is a specialized open-weights model by Google DeepMind. It is specifically built on top of the Gemma 3 architecture (e.g., the 270M parameter size) but tuned explicitly for function calling.

Key Features

On-Device Agents: Because of its tiny parameter size (270M), it is designed to run completely locally, on-device (e.g., phones or edge devices), serving as a fast, private agent.
Natural Language to API: It acts as a highly optimized base model for further training into custom agents that translate natural language commands directly into executable API actions.
Text-Only Focus: While the base Gemma 3 model introduces multimodal vision abilities, FunctionGemma is stripped down to focus purely on executing text-based function calling with an altered chat format.

TODO: Add prompt structure format for FunctionGemma.