This project provides a local Large Language Model (LLM) integration for Home Assistant, enabling users to control their smart home devices via natural language. It targets Home Assistant users seeking a private, offline conversational assistant for home automation. The primary benefit is enabling voice control and complex automation sequences without relying on cloud services.

How It Works

The solution consists of a custom Home Assistant conversation agent and fine-tuned LLMs. The agent exposes local LLMs (via llama-cpp-python, Ollama, or text-generation-webui) as a conversation backend within Home Assistant. The "Home" LLMs are fine-tuned on a synthetic dataset designed for function calling, allowing them to interpret natural language commands and map them to Home Assistant services (e.g., light.turn_on). The models are quantized for low-resource environments like Raspberry Pis.

Quick Start & Requirements

• Installation: Requires Home Assistant 2025.4.1 or newer. Install the custom component via Home Assistant's UI.
• Model Integration: Supports running models directly via llama-cpp-python within Home Assistant or remotely via Ollama, LocalAI, or text-generation-webui.
• Hardware: Quantized models are designed for low-resource environments (e.g., Raspberry Pi), but performance will vary.
• Documentation: Setup Guide

Highlighted Details

• Fine-tuned models available in 1B and 3B parameter sizes (e.g., acon96/Home-3B-v3-GGUF, acon96/Home-1B-v3-GGUF).
• Supports multiple prompt formats (Zephyr, ChatML) and experimental multilingual models.
• Includes an experimental Home Assistant Add-on for running text-generation-webui.
• Models are quantized using Llama.cpp for efficient execution.

Maintenance & Community

• Active development with frequent updates (see Version History).
• Supports a range of Home Assistant entity types (light, fan, cover, lock, media_player, climate, switch).
• Dataset available on HuggingFace: acon96/Home-Assistant-Requests

Licensing & Compatibility

• The project itself appears to be under a permissive license (likely MIT, based on common practice for such integrations, though not explicitly stated in the README).
• Models are hosted on HuggingFace, and their specific licenses should be checked individually.
• Compatible with Home Assistant's ecosystem.

Limitations & Caveats

The project is primarily focused on English, with experimental support for other languages. The "Home" models are fine-tuned for specific tasks and may have limitations in general QA or complex reasoning outside of smart home control. The minimum Home Assistant version requirement is quite recent (2025.4.1).