JackJ1/functiongemma-270m-it-mobile-actions-litertlm

📱 FunctionGemma-270M-IT: Mobile Actions (LiteRT)

This model is a specialized fine-tune of Google's FunctionGemma-270M-IT, specifically optimized for Mobile Actions. It is packaged in the .litertlm format, making it ready for immediate deployment on mobile devices using LiteRT (formerly TensorFlow Lite).

🚀 Deployment & Usage

1. Direct Use on Android

You can deploy this model directly using the Google AI Edge Gallery App.

1. Download the model file: mobile-actions_q8_ekv1024.litertlm
2. Open the Google AI Edge Gallery App: Navigate to the "Mobile Actions" section.
3. Load the Custom Model: Point the app to your downloaded .litertlm file.

2. Prompt Format

FunctionGemma requires an Essential System Prompt to activate its function-calling logic.

System Prompt (Developer Role):

You are a model that can do function calling with the following functions

Example User Input:

"Create a calendar event for lunch tomorrow at 12 PM"

Model Output:

<start_function_call>call:create_calendar_event{title:"Lunch", start_time:"2024-xx-xxT12:00:00"}<end_function_call>

---

❤️ Support & Community

If you find this model useful for your mobile AI projects, please consider leaving a Like ❤️ on the Hugging Face repository!

Special thanks to Google for the Gemma 3 architecture and the open-source community for making edge AI accessible.

---

Disclaimer: This model is designed for specialized function calling. It is not intended for general-purpose chatbot dialogue.

🌟 Key Highlights

• Ultra-Lightweight: At just 270M parameters, it's designed for low-latency, on-device inference without needing cloud connectivity.
• Expert at Mobile Tasks: Translates natural language (e.g., "Turn on the flashlight", "Set an alarm for 7 AM") into structured function calls for Android OS tools.
• Privacy Centric: Runs entirely offline, ensuring user queries and data remain private and secure on the device.
• Optimized Format: Provided as a quantized q8 .litertlm file for the best balance between performance and accuracy on edge hardware.

---

🛠 Model Technical Details

Detail	Specification
Base Model	google/functiongemma-270m-it
Architecture	Gemma 3 (270M)
Input Content	32K Tokens
Quantization	Q8 (8-bit)
Format	.litertlm (LiteRT)
Primary Task	Function Calling / Mobile Actions

---

📖 Background & Dataset

This model was fine-tuned using the Mobile Actions dataset and the official Google fine-tuning recipe.

It demonstrates how a small, specialized model can achieve state-of-the-art performance on specific agentic workflows, rivaling much larger models in specialized domains.

---