FunctionGemma

2025-12-18 · Publisher : Google · Domain : Artificial Intelligence

Google releases FunctionGemma, a version of Gemma 3 270M specialized in function calling, with its training recipe, to build local and private agents that translate natural language into API calls.

What changes

A 270-million-parameter model tuned to produce structured function calls and then summarize the results in natural language.
Can run as a standalone offline agent or serve as a router to larger models (Gemma 3 27B).
On the “Mobile Actions” evaluation, fine-tuning raises accuracy from 58% to 85%, according to Google.
A 256,000-token vocabulary that efficiently tokenizes JSON and multilingual inputs.

Google releases FunctionGemma, a version of Gemma 3 270M specialized in function calling, with its training recipe, to build local and private agents that translate natural language into API calls. ### What changes - A 270-million-parameter model tuned to produce structured function calls and then summarize the results in natural language. - Can run as a standalone offline agent or serve as a router to larger models (Gemma 3 27B). - On the “Mobile Actions” evaluation, fine-tuning raises accuracy from 58% to 85%, according to Google. - A 256,000-token vocabulary that efficiently tokenizes JSON and multilingual inputs. ### Specifications - Small size: runs on mobile and NVIDIA Jetson Nano. - Tools: Hugging Face Transformers, Unsloth, Keras, NVIDIA NeMo; deployment with LiteRT-LM, vLLM, MLX, llama.cpp, Ollama, Vertex AI, LM Studio. - Availability: Hugging Face and Kaggle. ### Pricing Open model, with no usage pricing of its own. ### Benchmarks | Evaluation | Score | |---|---| | Mobile Actions, baseline | 58% | | Mobile Actions, after fine-tuning | 85% | Scores published by the vendor. ### Key takeaways A building block for privacy-respecting embedded agents. It requires fine-tuning to reach good reliability.

FunctionGemma

What changes

Links