your-model-man/qwen3.5-4b-home-assistant

Qwen3.5-4B Home Assistant

A fine-tuned version of Qwen3.5-4B trained to control Home Assistant via natural language. Given a user command, the model outputs a structured JSON tool call that Home Assistant can execute directly through its OpenAI-compatible conversation API.

Fine-tuned using Unsloth for 2x faster training with LoRA adapters, then exported to GGUF Q4_K_M for efficient local inference.

Model Details

Property	Value
Base model	Qwen3.5-4B
Fine-tuning method	LoRA (r=8, alpha=8)
Training framework	Unsloth
Training examples	810
Epochs	3
Final training loss	~0.023
Quantization	Q4_K_M GGUF
Context length	1024 tokens

Intended Use

Designed to be served locally via llama.cpp and connected to Home Assistant as a local conversation agent — no cloud, no API keys, full privacy.

Supported device domains:

• Lights (on/off, brightness, colour, colour temperature)
• Climate / thermostat (get/set temperature)
• Covers & blinds (open/close/position)
• Fans (on/off, speed)
• Locks (lock/unlock, state query)
• Media players (play/pause/next/previous/volume/mute)
• Vacuum cleaners (start/dock/clean area)
• Shopping & todo lists
• Scenes & scripts
• Timers
• Weather queries
• Date & time queries

Output Format

For device control commands the model outputs a JSON tool call:

{"tool": "HassTurnOn", "parameters": {"name": "kitchen light", "domain": "light"}}

{"tool": "HassLightSet", "parameters": {"area": "bedroom", "brightness": 50}}

{"tool": "HassClimateSetTemperature", "parameters": {"temperature": 72}}

For conversational messages it responds in plain text.

How to Use

With llama.cpp server

llama-server \
  --model Qwen3.5-4B.Q4_K_M.gguf \
  --port 8080 \
  --host 0.0.0.0 \
  --ctx-size 1024 \
  --chat-template chatml

Connect to Home Assistant

1. Go to Settings → Devices & Services → Add Integration → OpenAI Conversation
2. Set Base URL to http://your-local-ip:8080/v1
3. Set API key to any value (e.g. sk-local)
4. Assign it as your conversation or voice agent

In Python

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="luc-h/qwen3.5-4b-home-assistant",
    filename="Qwen3.5-4B.Q4_K_M.gguf",
)

response = llm.create_chat_completion(messages=[
    {"role": "system", "content": "You are a Home Assistant AI. Respond with JSON tool calls for device control."},
    {"role": "user", "content": "turn off the kitchen lights"},
])
print(response["choices"][0]["message"]["content"])
# {"tool": "HassTurnOff", "parameters": {"area": "kitchen", "domain": "light"}}

Training Data

Generated from the official Home Assistant intents repository — the same sentence templates used by Home Assistant's built-in Assist voice assistant. Examples cover all major device domains with varied natural language phrasings, entity names, and area references.

Limitations

• Entity and area names in responses are placeholders — Home Assistant's conversation pipeline matches these to your actual devices
• Fine-tuned on English only
• Not designed for complex multi-step automations or scripting

Training Hardware

Trained on Google Colab free tier (NVIDIA T4, 15GB VRAM) in approximately 1 hour.

License

Apache 2.0 — same as the base Qwen3.5 model.