home-llm

Home LLM

Control your Home Assistant smart home with a completely local Large Language Model. No cloud services and no subscriptions needed. Just privacy-focused AI running entirely on your own hardware.

What is Home LLM?

Home LLM is a complete solution for adding AI-powered voice and chat control to Home Assistant. It consists of two parts:

1. Local LLM Integration – A Home Assistant custom component that connects local language models to your smart home
2. Home Models – Small, efficient AI models fine-tuned specifically for smart home control

Key Features

• Fully Local – Everything runs on your hardware. Your data never leaves your control (unless you want to!)
• ️ Voice & Chat Control – Use as a conversation agent with voice assistants or chat interfaces
• AI Task Automation – Generate dynamic content and structured data for automations
• Multi-Language Support – Built-in support for English, German, French, Spanish, and Polish (better translations are welcome!)
• ⚡ Runs on Low-Power Devices – Models work on Raspberry Pi and other modest hardware -- no GPU required!
• Flexible Backends – Run models locally as part of Home Assistant or connect to external model providers

Quick Start

See the Setup Guide for detailed installation instructions.

Requirements: Home Assistant 2026.5.0 or newer

---

Local LLM Integration

The integration connects language models to Home Assistant, enabling them to understand your requests and control your smart devices.

Supported Backends

Choose how and where you want to run your models:

Backend	Best For
Llama.cpp (built-in)	Running models directly in Home Assistant
Ollama	Easy setup on a separate GPU machine
Generic OpenAI API	LM Studio, LocalAI, vLLM, and other OpenAI-compatible servers
llama.cpp server	Heterogeneous (non-uniform) GPU compute setups, including CPU + GPU inference
OpenAI 'Responses' Style API	Cloud services supporting the 'responses' style API
Anthropic 'Messages' Style API	Cloud services supporting the 'messages' style API
text-generation-webui	Advanced users with existing setups

NOTE: When utilizing external APIs or model providers, your data will be transmitted over the internet and shared with the respective service providers. Ensure you understand the privacy implications of using these third-party services, since they will be able to see the status of all exposed entities in your Home Assistant instance, which can potentially include your current location.

Supported Device Types

The integration can control: lights, switches, fans, covers, locks, climate, media players, vacuums, buttons, timers, todo lists, and scripts

Using the Integration

As a Conversation Agent:

• Chat with your assistant through the Home Assistant UI
• Connect to voice pipelines with Speech-to-Text and Text-to-Speech
• Supports voice streaming for faster responses

As an AI Task Handler:

• Create automations that use AI to process data and generate structured responses
• Perfect for dynamic content generation, data extraction, and intelligent decision making
• See AI Tasks documentation for examples

---

Home LLM Models

The "Home" models are small language models (under 5B parameters) fine-tuned specifically for smart home control. They understand natural language commands and translate them into Home Assistant service calls.

Latest Models

Model Family	Size	Link
Llama 3.2	3B	acon96/Home-Llama-3.2-3B
Gemma	270M	acon96/Home-FunctionGemma-270m

Previous Model Versions

Stable Models:

• 3B v3 (StableLM-Zephyr-3B): acon96/Home-3B-v3-GGUF
• 1B v3 (TinyLlama-1.1B): acon96/Home-1B-v3-GGUF
• 3B v2 (Phi-2): acon96/Home-3B-v2-GGUF
• 1B v2 (Phi-1.5): acon96/Home-1B-v2-GGUF
• 1B v1 (Phi-1.5): acon96/Home-1B-v1-GGUF

Multilingual Experiments:

• German, French, & Spanish (3B): acon96/stablehome-multilingual-experimental
• Polish (1B): acon96/tinyhome-polish-experimental

Note: Models v1 (3B) and earlier are only compatible with integration version 0.2.17 and older.

Using Other Models

Don't have dedicated hardware? You can use any instruction-tuned model with in-context learning (ICL). The integration provides examples that teach general-purpose models (like Qwen3, Llama 3, Mistral) how to control your smart home. See the Setup Guide for configuration details.

Training Your Own

The fine-tuning dataset and training scripts are included in this repository:

• Dataset: Home-Assistant-Requests-V2 on HuggingFace
• Source: data/ directory
• Training: See train/README.md

---

Documentation

• Setup Guide – Installation and configuration
• Backend Configuration – Detailed backend options
• Model Prompting – Customize system prompts
• AI Tasks – Using AI in automations

---

Version History

Version	Highlights
v0.4.11	Compatibility fix for Home Assistant 2026.9.0 and newer
v0.4.10	Bug fixes and manual parsing improvements for tool calls and thinking blocks
v0.4.9	Relax dependency requirements to avoid conflicting with internal HA lib versions
v0.4.8	OpenAI backends rewritten using the official openai Python library for better reliability and compatibility. New "Use server sampling defaults" to let your backend set the sampling parameters. More robust tool call parsing with auto-repair for malformed JSON, ability to disable streaming for all backends.
v0.4.7	Bug fixes, update default llama_cpp_python version to support new models, and support python 3.14 for new Home Assistant versions
v0.4.6	Anthropic API support, on-disk caching for Llama.cpp, new tool calling dataset
v0.4.5	AI Task entities, multiple LLM APIs at once, official Ollama package
v0.4	Tool calling rewrite, voice streaming, agentic tool use loop, multiple configs per backend
v0.3	Home Assistant LLM API support, improved prompting, HuggingFace GGUF auto-detection

Full Version History

Version	Description
v0.4.11	Compatibility fix for Home Assistant 2026.9.0 and newer
v0.4.10	Bug fixes and manual parsing improvements for tool calls and thinking blocks
v0.4.9	Relax dependency requirements to avoid