Distil-Home-Assistant-Qwen3

A fine-tuned Qwen3-0.6B model for multi-turn intent classification and slot extraction in an on-device smart home controller. Trained using knowledge distillation from a 120B teacher model, this 0.6B model delivers 96.7% tool call accuracy — exceeding the teacher — while running locally on-device for private, low-latency smart home control.

For the GGUF version (for llama.cpp / Ollama deployment), see distil-labs/distil-home-assistant-qwen3-gguf.

Results

Model	Parameters	Tool Call Accuracy	ROUGE
GPT-oss-120B (teacher)	120B	94.1%	98.2%
This model (tuned)	0.6B	96.7%	99.2%
Qwen3-0.6B (base)	0.6B	—	—

The fine-tuned model exceeds the 120B teacher on tool call accuracy while being 200x smaller. Fine-tuning is essential for reliable multi-turn smart home conversations.

Quick Start

Using Transformers

import json
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("distil-labs/distil-home-assistant-qwen3")
tokenizer = AutoTokenizer.from_pretrained("distil-labs/distil-home-assistant-qwen3")

TOOLS = [
    {"type": "function", "function": {"name": "toggle_lights", "description": "Turn lights on or off in a specified room", "parameters": {"type": "object", "properties": {"room": {"type": "string", "enum": ["living_room", "bedroom", "kitchen", "bathroom", "office", "hallway"]}, "state": {"type": "string", "enum": ["on", "off"]}}, "required": [], "additionalProperties": False}}},
    {"type": "function", "function": {"name": "set_thermostat", "description": "Set the temperature for heating or cooling", "parameters": {"type": "object", "properties": {"temperature": {"type": "integer", "minimum": 60, "maximum": 80}, "mode": {"type": "string", "enum": ["heat", "cool", "auto"]}}, "required": [], "additionalProperties": False}}},
    {"type": "function", "function": {"name": "lock_door", "description": "Lock or unlock a door", "parameters": {"type": "object", "properties": {"door": {"type": "string", "enum": ["front", "back", "garage", "side"]}, "state": {"type": "string", "enum": ["lock", "unlock"]}}, "required": [], "additionalProperties": False}}},
    {"type": "function", "function": {"name": "get_device_status", "description": "Query the current state of a device or room", "parameters": {"type": "object", "properties": {"device_type": {"type": "string", "enum": ["lights", "thermostat", "door", "all"]}, "room": {"type": "string"}}, "required": [], "additionalProperties": False}}},
    {"type": "function", "function": {"name": "set_scene", "description": "Activate a predefined scene", "parameters": {"type": "object", "properties": {"scene": {"type": "string", "enum": ["movie_night", "bedtime", "morning", "away", "party"]}}, "required": [], "additionalProperties": False}}},
    {"type": "function", "function": {"name": "intent_unclear", "description": "Use when the user's intent cannot be determined", "parameters": {"type": "object", "properties": {"reason": {"type": "string", "enum": ["ambiguous", "off_topic", "incomplete", "unsupported_device"]}}, "required": [], "additionalProperties": False}}},
]

messages = [
    {"role": "system", "content": "You are a tool-calling model working on:\n<task_description>You are an on-device smart home controller. Given a natural language command from the user, call the appropriate smart home function. If the user does not specify a required value (e.g. which room or what temperature), omit that parameter from the function call. Maintain context across conversation turns to resolve pronouns and sequential commands.</task_description>\n\nRespond to the conversation history by generating an appropriate tool call that satisfies the user request. Generate only the tool call according to the provided tool schema, do not generate anything else. Always respond with a tool call.\n\n"},
    {"role": "user", "content": "Turn off the living room lights"},
]

text = tokenizer.apply_chat_template(
    messages, tools=TOOLS, tokenize=False, add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
# <tool_call>
# {"name": "toggle_lights", "arguments": {"room": "living_room", "state": "off"}}
# </tool_call>

Using Ollama

huggingface-cli download distil-labs/distil-home-assistant-qwen3 --local-dir distil-model
cd distil-model
ollama create distil-home-assistant -f Modelfile
ollama run distil-home-assistant

Using with the Demo App

This model powers the Smart Home Controller demo — a text-based orchestrator that pairs an SLM with deterministic dialogue management for smart home control.

Model Details

Property	Value
Base Model	Qwen/Qwen3-0.6B
Parameters	0.6 billion
Architecture	Qwen3ForCausalLM
Context Length	40,960 tokens
Precision	bfloat16
Training Data	50 seed conversations, synthetically expanded
Teacher Model	GPT-oss-120B
Task	Multi-turn tool calling (closed book)

Training

This model was trained using the Distil Labs platform:

Seed Data: 50 hand-written multi-turn conversations covering 6 smart home functions with 2-5 user turns per conversation
Synthetic Expansion: Expanded to thousands of examples using a 120B teacher model
Fine-tuning: Multi-turn tool calling distillation on Qwen3-0.6B

What the Model Does

The model acts as a function caller for a smart home controller. Given a user command and conversation history, it outputs a structured tool call:

User: "Turn off the living room lights"
Model: {"name": "toggle_lights", "arguments": {"room": "living_room", "state": "off"}}

User: "The kitchen too"
Model: {"name": "toggle_lights", "arguments": {"room": "kitchen", "state": "off"}}

User: "Lock the front door"
Model: {"name": "lock_door", "arguments": {"door": "front", "state": "lock"}}

User: "Can you order me a pizza?"
Model: {"name": "intent_unclear", "arguments": {"reason": "off_topic"}}

Supported Functions

The model handles 6 smart home operations:

Function	Description
`toggle_lights`	Turn lights on or off in a room
`set_thermostat`	Set temperature and heating/cooling mode
`lock_door`	Lock or unlock a door
`get_device_status`	Query device state
`set_scene`	Activate a predefined scene
`intent_unclear`	Cannot determine intent

Use Cases

On-device smart home controllers with privacy-first design
Text-based smart home chatbots with structured intent routing
Edge deployment for local smart home hubs
Any multi-turn tool calling task with bounded intent taxonomy

Limitations

Trained on English smart home intents only
Covers 6 specific smart home functions — not a general-purpose tool caller
96.7% accuracy means a small fraction of function calls may be incorrect
Temperature range is fixed to 60-80°F

License

This model is released under the Apache 2.0 license.

Citation

@misc{distil-home-assistant-qwen3,
  author = {Distil Labs},
  title = {Distil-Home-Assistant-Qwen3: A Fine-tuned SLM for Smart Home Control},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/distil-labs/distil-home-assistant-qwen3}
}

Downloads last month: 28

Safetensors

Model size

0.6B params

Tensor type

BF16

Model tree for distil-labs/distil-home-assistant-qwen3

Base model

Qwen/Qwen3-0.6B-Base

Finetuned

Qwen/Qwen3-0.6B