As the world pays attention to the growth of larger and larger models, Google released something incredibly small and potentially even more transformative for everyday developers.
Google released FunctionGemma, a 270M parameter model that doesn’t chat, it act. FunctionGemma is a lightweight, open model developed based on Gemma 3 architecture and tailored for function calls and agentic workflows at the edge.
It doesn’t write poetry or solve philosophy; rather it translates natural language input to a structured format (such as JSON or API calls) necessary to operate software and hardware in real time.
The Intelligent Traffic Controller
Performance is the cornerstone of FunctionGemma. Google calls it an “intelligent traffic controller” for AI systems.
Rather than passing every basic request (e.g. “turn on the lights” or ”set a timer”) to an enormous, resource-hungry cloud model, FunctionGemma is located natively on your device (including smartphones, laptops and NVIDIA Jetson Nano).
It does these basics things quickly and without a connection. If there’s a request it deems too complicated for its 270M brain, it can route the user to another model like Gemma 3 27B or Gemini.
Also Read: The ChatGPT Store Isn’t Just a Launch, It’s the Beginning of an AI App Economy
Small but Deterministic
When it comes to the world of function calling, predictability trumps creativity. FunctionGemma is designed to be deterministic (i.e. it returns the correct code structure and does not have “hallucinations” as chat models often can), fine-tuned on behavior.
- Performance: The base model obtains a 58% accuracy on the “Mobile Actions” dataset, right after initialization. But with focused fine-tuning, this goes up to 85%, which is sufficient for production quality mobile agents.
- Hardware Friendly: With its small size, it can run on commodity hardware with super low-latency. It can respond to requests, on a Pixel 8 or iPhone 15 Pro at about ~50 tokens per second.
Demos: Small Gardens And Physics Puzzles
To demonstrate the model, Google released some interactive demos:
- TinyGarden: A voice-activated game in which people run a virtual smallholding. Instructions such as “plant sunflowers on the top row” are immediately parsed into game logic, without touching a server.
- Mobile Actions: An offline-only example implementation that demonstrates how your model can manage Android OS features (calendar, flashlight, contacts).
With support from major developer tools like Unsloth, Ollama, and Hugging Face from day one, FunctionGemma is ready to be dropped into applications immediately.
