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

FunctionGemma: Enabling On-Device AI Agents

FunctionGemma is a new, lightweight iteration of the Gemma 3 270M model specifically engineered for translating natural language into structured function and API calls. This allows AI agents to move beyond simple conversation and actively perform tasks, like scheduling events or controlling device settings.

The ideal of a general-purpose LLM often clashes with the realities of resource-constrained environments; large models struggle with latency and power consumption on edge devices. FunctionGemma addresses this by prioritizing specialization and efficient execution, enabling practical AI applications directly on mobile phones and similar hardware.

Key Insights

• Mobile Actions Accuracy: Fine-tuning improved FunctionGemma’s reliability from 58% to 85% on mobile action tasks (Google, 2026).
• Unified Action & Chat: The model seamlessly switches between generating structured function calls and providing natural language explanations.
• Framework Support: FunctionGemma is compatible with Hugging Face Transformers, Unsloth, Keras, NVIDIA NeMo, vLLM, MLX, Llama.cpp, Ollama, Vertex AI, and LM Studio.

Working Example

# Example: Physics Playground interaction (simplified)
# Using Transformer.js for client-side JavaScript integration
# Natural language input: "Increase gravity slightly."

# FunctionGemma processes the input and generates a function call:
function_call = {
  "function": "set_gravity",
  "parameters": {
    "value": 1.1 # Example: Increase gravity by 10%
  }
}

# The application executes the function call:
simulation.set_gravity(function_call["parameters"]["value"])

# FunctionGemma then generates a natural language response:
response = "Gravity has been increased slightly."

Practical Applications

• Mobile Assistants: Google’s Mobile Actions demo demonstrates automating tasks like creating calendar events or controlling phone settings.
• Pitfall: Relying on zero-shot prompting with a general-purpose LLM for function calling often results in unreliable and unpredictable behavior, requiring substantial fine-tuning for production use.

References:

• https://www.infoq.com/news/2026/01/functiongemma-edge-function-call/