A
argbe.tech - news
1min read

FunctionGemma targets reliable tool calling with a Gemma 3 270M fine-tuned core

Google DeepMind introduced FunctionGemma, a Gemma 3 270M-derived model tuned to map natural-language requests into executable tool calls. A no-training-code Tuning Lab and common open-source workflows aim to make tool routing easier to customize.

Google DeepMind released FunctionGemma, a specialized Gemma 3 270M fine-tune designed to turn natural language into executable software actions for tool-calling agents.

  • FunctionGemma Tuning Lab supports tool-calling fine-tunes without writing training code.
  • Fine-tuning is positioned as a way to reduce confusion when choosing between similar tools (for example, an internal knowledge base vs. a public web search).
  • Hugging Face TRL is highlighted as the main library used in practical fine-tuning workflows.
  • The bebechien/SimpleToolCalling dataset is referenced for conversational examples that train and evaluate tool-routing behavior.
  • Distillation is presented as an option for training with synthetic data generated by larger models.