🚀 An experiment, and an open invitation. Domux explores a new idea: how far can text semantic parsing go under an aggressive latency budget — keeping end-to-end response under 150ms? This is an early-stage exploration, and we're sharing it in the hope that others will try it too. If this direction interests you, we'd love for you to follow along and explore it together.
Domux (Domux-Gemma-4-E2B-it) is a fine-tuned language model built on Gemma-4-E2B-it. It turns natural-language smart-home commands into structured, pipe-delimited slots. Training combines supervised fine-tuning (SFT) with reinforcement learning via Group Relative Policy Optimization (GRPO) and custom reward functions.
- 2026.06.30 — 🎉 Released v0.1.0, the first open-source release. See CHANGELOG.
- 2026.06.29 — Released training code, reward plugins, and example datasets.
- 2026.06.25 — Initial release of Domux based on Gemma-4-E2B-it.
Full version history: CHANGELOG.md
- Fast response — Optimized for low-latency inference on edge devices and servers.
- Structured slot output — Parses free-form commands into a fixed 7-field pipe-delimited schema.
- High accuracy — 98.37% result accuracy with 100% format compliance, outperforming much larger models.
- Lightweight base — Built on the compact Gemma-4-E2B-it, suitable for on-device and edge deployment.
- Multi-action support — Handles compound commands that map to multiple slot lines.
- Generalizes across devices — Handles arbitrary device names within each category, not a fixed whitelist.
Domux is built on a generalist base model that does not rely on a fixed device whitelist — the model handles diverse device names through semantic understanding. The tables below show core capabilities; full specification at Output Format Documentation.
| Device Type | Naming Examples | Controllable Attributes | Value Range |
|---|---|---|---|
| Light | Light、Strip Light、Spot Light、Desk Lamp | brightnesscolorcolorTemperaturemode |
0–100% Blue、Red、Green、Yellow、White, etc. 3000–6500 K Reading、Romance、Soft, etc. |
| AC | AC、AC 1 | temperaturemodewindSpeed |
16–30 °C Cool、Heat、Dry、Fan、Auto Low、Medium、High |
| Curtain / Blind | Curtain、Blind、Sheer Curtain | position |
0–100% |
| Scene Mode | Romantic Mode、Party Mode、Sleeping Mode | — | — |
| Action | Description | Example Commands |
|---|---|---|
turnOn |
Turn on device | "Turn on the living room light" |
turnOff |
Turn off device | "Turn off the AC" |
set |
Set to specific value | "Set brightness to 80%"、"Set AC to 24 degrees" |
adjustUp |
Increase attribute | "Make it brighter"、"Increase the temperature a bit" |
adjustDown |
Decrease attribute | "Dim the light"、"Turn down the AC" |
activate |
Activate scene mode | "Activate romantic mode" |
deactivate |
Deactivate scene mode | "Deactivate party mode" |
pause |
Pause curtain movement | "Stop the curtain" |
Room Support: Living Room、Bedroom、Kitchen、Bathroom、Home Office、Balcony、Master Bedroom, etc. Supports numbering (Bedroom 1、Room A) and culturally specific rooms (Majlis、Prayer Room).
Floor Support: Ground Floor、First Floor、Second Floor、Upstairs、Downstairs, etc.
Adjustment commands without explicit values (e.g., "make it brighter", "turn it down a bit") map to adjustUp / adjustDown with the value field left empty (*), allowing downstream systems to determine the adjustment magnitude based on current state.
As an early exploration, Domux is still evolving. We're focusing on three directions:
- 🔜 Broader device coverage — more categories beyond lighting, climate, window, and audio
- 🔜 Richer scenes — support for more scenes and modes
- 🔜 Stronger fuzzy-intent understanding — better handling of vague, implicit, and context-dependent commands
The model outputs pipe-delimited slots with 7 fields:
action|device|attribute|value|unit|room|floor
| Input | Output |
|---|---|
| Turn on the living room light | turnOn|Light|*|*|*|Living Room|* |
| Set bedroom AC to 22 degrees | set|AC|temperature|22|Celsius|Bedroom|* |
| Close the curtains 20 percent | adjustDown|Curtain|openness|20|Percent|*|* |
Input:
Turn on the Master Light in the Master Bedroom on the Second Floor,
set brightness to 80%, color temperature to 4000K, color to Blue, and mode to Reading.
Output:
turnOn|Light|*|*|*|Master Bedroom|Second Floor
set|Light|brightness|80|Percent|Master Bedroom|Second Floor
set|Light|colorTemperature|4000|Kelvin|Master Bedroom|Second Floor
set|Light|color|Blue|*|Master Bedroom|Second Floor
set|Light|mode|Reading|*|Master Bedroom|Second Floor
Input:
Turn off all lights in the Living Room on the Ground Floor,
set the AC to Cool mode at 24 degrees in the Guest Bedroom,
and open the curtains halfway in the Dining Room.
Output:
turnOff|Light|*|*|*|Living Room|Ground Floor
set|AC|mode|Cool|*|Guest Bedroom|*
set|AC|temperature|24|Celsius|Guest Bedroom|*
set|Curtain|openness|50|Percent|Dining Room|*
Use * for unspecified or don't-care fields.
Evaluated on a comprehensive test set of 4,057 samples across 4 dimensions (single intent, multi-intent, omitted attributes, non-standard naming), benchmarked against 11 mainstream models including Qwen3.5 series (2B-27B), Gemma 4 series, and leading closed-source APIs (DeepSeek-V4, Claude Haiku 4.5, Gemini 3.5 Flash).
📄 Full technical report: English Report ・ 中文报告
We've open-sourced the test set and evaluation script behind the benchmark above, so you can reproduce our results or evaluate your own model.
- Test set —
eval/smart_home_control_test_set.jsonl: 4,057 samples across four dimensions (single-intent, multi-intent, omitted attributes, non-standard naming), spanning three device categories (Light / AC / Curtain) with 67 device-name variants. - Evaluation script —
eval/run_eval.py: sends each query to any OpenAI-compatible endpoint and reports format compliance, result accuracy, Slot F1, Intent F1, and average latency. - Docs — see
eval/DATASET_README.mdfor data format, metric definitions, and how to run.
pip install requests
# Fill in API_KEY / BASE_URL / MODEL in run_eval.py, then run
python eval/run_eval.py| Model | Base | Hugging Face | ModelScope |
|---|---|---|---|
| Domux-Gemma-4-E2B-it | Gemma-4-E2B-it | 🤗 iFlytekOpenSource/Domux | 🔧 iflytek/domux |
Download via:
# Option 1: Hugging Face
git lfs install
git clone https://huggingface.co/iFlytekOpenSource/Domux
# Option 2: ModelScope
pip install modelscope
python -c "from modelscope import snapshot_download; snapshot_download('iflytek/domux', cache_dir='./models')"The model runs in BF16 precision and requires 20GB+ of VRAM for single-GPU deployment.
# Option 1: vLLM
pip install "vllm==0.22.0"
# Option 2: SGLang
pip install "sglang[all]==0.5.12"Offline inference with vLLM. Pass the user command directly as the query:
from vllm import LLM, SamplingParams
# Point to your downloaded model directory
llm = LLM(model="Domux", dtype="bfloat16") # or "models/iflytek/domux"
sampling = SamplingParams(temperature=0.0, max_tokens=256)
prompt = "Turn on the Master Light in the Master Bedroom on the Second Floor, set brightness to 80%, color temperature to 4000K, color to Blue, and mode to Reading."
output = llm.chat([{"role": "user", "content": prompt}], sampling)
print(output[0].outputs[0].text)
# Output:
# turnOn|Light|*|*|*|Master Bedroom|Second Floor
# set|Light|brightness|80|Percent|Master Bedroom|Second Floor
# set|Light|colorTemperature|4000|Kelvin|Master Bedroom|Second Floor
# set|Light|color|Blue|*|Master Bedroom|Second Floor
# set|Light|mode|Reading|*|Master Bedroom|Second FloorServe the model as an OpenAI-compatible API with either vLLM or SGLang.
python -m vllm.entrypoints.openai.api_server \
--model Domux \
--served-model-name domux \
--host 0.0.0.0 \
--port 8000 \
--dtype bfloat16 \
--max-model-len 2048 \
--gpu-memory-utilization 0.9python -m sglang.launch_server \
--model-path Domux \
--host 0.0.0.0 \
--port 8000 \
--dtype bfloat16 \
--context-length 2048Both servers expose the same OpenAI-compatible endpoint:
from openai import OpenAI
client = OpenAI(
base_url="http://localhost:8000/v1",
api_key="EMPTY"
)
response = client.chat.completions.create(
model="domux",
messages=[
{"role": "user", "content": "Turn on the Master Light in the Master Bedroom on the Second Floor, set brightness to 80%, color temperature to 4000K, color to Blue, and mode to Reading."}
],
temperature=0.0
)
res = response.choices[0].message.content
print(res)
# Output:
'''
turnOn|Light|*|*|*|Master Bedroom|Second Floor
set|Light|brightness|80|Percent|Master Bedroom|Second Floor
set|Light|colorTemperature|4000|Kelvin|Master Bedroom|Second Floor
set|Light|color|Blue|*|Master Bedroom|Second Floor
set|Light|mode|Reading|*|Master Bedroom|Second Floor
'''See the LICENSE file in this repository.
- Base model: Gemma
- Training framework: ModelScope-Swift
- Experiment tracking: SwanLab
Join the Astron Open Source Community (WeCom Group) to discuss and collaborate:
