Google Releases Gemma 3 270M Variant Optimized for Function Calling on Mobile and Edge Devices

"FunctionGemma is a new, lightweight version of the Gemma 3 270M model, fine-tuned to translate natural language into structured function and API calls, enabling AI agents to "do more than just talk" and act. Launched a few months ago, Gemma 3 270M has evolved into FunctionGemma by gaining native function call capabilities in response to developer demand. Running locally allows the model to function either as an independent agent for private, offline tasks or as an "intelligent traffic controller" that routes more complex requests to larger, remote models."

"This is particularly compelling on-device, where agents can automate complex, multi-step workflows, from setting reminders to toggling system settings. To enable this at the edge, models must be lightweight enough to run locally and specialized enough to be reliable. Rather than being intended for zero-shot prompting, FunctionGemma is designed to be further customized into fast, private, on-device agents capable of translating natural language into executable API actions, Google explains. This approach is key to achieving production-ready level performance."

"In our "Mobile Actions" evaluation, fine-tuning transformed the model's reliability, boosting accuracy from a 58% baseline to 85%. The model is engineered to run efficiently on resource-constrained devices like mobile phones and NVIDIA Jetson Nano. It uses Gemma's 256k vocabulary to tokenize JSON and multilingual inputs efficiently. FunctionGemma supports what Google calls "unified action and chat", enabling it to generate structured code/function calls to execute tools and then switch back to natural language to explain the results to the user. Google also notes that FunctionGemma offers broad ecosystem support, allowing fine-tuning with frameworks such as Hugging Face Transformers, Unsloth, Keras or NVIDIA NeMo and deployment via LiteRT-LM, vLLM, MLX, Llama.cpp, Ollama, Vertex AI or LM"

FunctionGemma originates from Gemma 3 270M and adds native function-call capabilities to convert natural language into structured API and function calls. The model runs locally to operate as private, offline agents or to route complex requests to larger remote models. Fine-tuning focuses on producing fast, private, on-device agents rather than zero-shot prompting, improving task reliability in multi-step workflows. Evaluations show fine-tuning raised Mobile Actions accuracy from 58% to 85%. The model is optimized for resource-constrained hardware, uses a 256k vocabulary for efficient JSON and multilingual tokenization, and supports unified action-and-chat interactions plus broad ecosystem fine-tuning and deployment options.

Read at InfoQ