distil-labs/distil-home-assistant-functiongemma
Distil-Home-Assistant-FunctionGemma by Distil Labs is a 0.3 billion parameter FunctionGemma (Gemma3 architecture) model fine-tuned for on-device smart home control. It excels at multi-turn intent classification and slot extraction, achieving 96.71% tool call accuracy through knowledge distillation from a 120B teacher model. This specialized model is designed for private, low-latency local execution in smart home environments.
Loading preview...
Overview
Distil-Home-Assistant-FunctionGemma is a compact, 0.3 billion parameter model from Distil Labs, built on the FunctionGemma (Gemma3) architecture. It is specifically fine-tuned for on-device smart home control, focusing on multi-turn intent classification and slot extraction. This model achieves an impressive 96.71% tool call accuracy, surpassing its 120B parameter teacher model, making it highly reliable for smart home automation.
Key Capabilities
- High Accuracy: Achieves 96.71% tool call accuracy, outperforming the base FunctionGemma and even the larger 120B teacher model.
- On-Device Performance: Designed for local execution, ensuring privacy and low-latency responses for smart home commands.
- Multi-Turn Tool Calling: Capable of maintaining context across conversations to resolve pronouns and sequential commands for 6 specific smart home functions (e.g.,
toggle_lights,set_thermostat,lock_door). - Knowledge Distillation: Trained using a synthetic dataset expanded from 50 seed conversations, distilled from a 120B teacher model.
Use Cases
- Privacy-First Smart Home Controllers: Ideal for local, privacy-focused smart home hubs.
- Edge Deployment: Suitable for deployment on resource-constrained edge devices.
- Text-Based Automation: Powers text-based smart home chatbots with structured intent routing.
- Bounded Tool Calling: Effective for multi-turn tool calling tasks within a defined set of intents, such as smart home operations.