This research demonstrates how fine-tuned LLMs can run directly on resource-limited home devices to both understand user commands and generate natural responses.

• Successfully trained models to produce JSON action calls and human-like text responses
• Achieved high accuracy using both 16-bit and 8-bit quantized models on CPU-only hardware
• Eliminated cloud dependency for smart home commands, significantly enhancing privacy and security
• Proved that even resource-constrained devices can run sophisticated LLMs for home automation

Security Impact: By processing all data locally rather than in the cloud, this approach drastically reduces privacy risks associated with smart home systems, eliminating potential data interception and unauthorized access concerns.

On-Device LLMs for Home Assistant: Dual Role in Intent Detection and Response Generation