Instructions to use henrywangxf/lux with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use henrywangxf/lux with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="henrywangxf/lux",
	filename="qwen3-1.7b.Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps

llama.cpp

How to use henrywangxf/lux with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf henrywangxf/lux:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf henrywangxf/lux:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf henrywangxf/lux:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf henrywangxf/lux:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf henrywangxf/lux:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf henrywangxf/lux:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf henrywangxf/lux:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf henrywangxf/lux:Q4_K_M

Use Docker

docker model run hf.co/henrywangxf/lux:Q4_K_M

LM Studio
Jan

vLLM

How to use henrywangxf/lux with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "henrywangxf/lux"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "henrywangxf/lux",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/henrywangxf/lux:Q4_K_M

Ollama
How to use henrywangxf/lux with Ollama:
```
ollama run hf.co/henrywangxf/lux:Q4_K_M
```

Unsloth Studio new

How to use henrywangxf/lux with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for henrywangxf/lux to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for henrywangxf/lux to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for henrywangxf/lux to start chatting

Pi new

How to use henrywangxf/lux with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf henrywangxf/lux:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "lux"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use henrywangxf/lux with Docker Model Runner:
```
docker model run hf.co/henrywangxf/lux:Q4_K_M
```

Lemonade

How to use henrywangxf/lux with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull henrywangxf/lux:Q4_K_M

Run and chat with the model

lemonade run user.lux-Q4_K_M

List all available models

lemonade list

lux

Fine-tuned Qwen3-1.7B for Linux system administration tool calls. Powers the slow path of the lux CLI — when the rule-based fast path can't classify a request, this model parses it into a structured tool invocation.

This repo ships the raw GGUF weights + Modelfile. If you want the full CLI experience (with luxd, ollama, pre-pulled weights, and a REPL), use henrywangxf/lux-portable instead.

Usage

With ollama (recommended)

ollama pull hf.co/henrywangxf/lux
ollama run hf.co/henrywangxf/lux "restart nginx"

Ollama reads the bundled Modelfile automatically — SYSTEM prompt, template, and sampling params come with the pull.

With llama.cpp

wget https://huggingface.co/henrywangxf/lux/resolve/main/qwen3-1.7b.Q4_K_M.gguf
./llama-cli -m qwen3-1.7b.Q4_K_M.gguf -p "restart nginx"

You'll need to supply your own prompt template — see Modelfile in this repo for the one lux expects.

What the model outputs

Trained to emit tool calls for lux's registry. Typical interaction:

user:    my disk is almost full
model:   {"tool": "check_disk_usage", "args": {}}

user:    install firefox
model:   {"tool": "install_flatpak", "args": {"app": "org.mozilla.firefox"}}

The full tool schema lives in lux-tools. If you're using this model outside of lux, you'll need to parse these tool calls and dispatch them yourself.

Training

Base: Qwen/Qwen3-1.7B (Apache 2.0)
Method: LoRA fine-tuning on a curated set of Linux sysadmin tool-use traces
Quantization: Q4_K_M (GGUF)
Recipe: see finetune/ in the main repo

Intended use

Parsing natural-language sysadmin requests into structured tool calls
Running locally on a laptop (≥8 GB RAM is comfortable for Q4_K_M)
Integrating with your own agent loop over lux's tool schema — or using lux directly

Limitations

English only.
Tuned for Fedora / RHEL-family systems (dnf, systemd, firewalld). Other distros work for most tools but package-manager queries degrade.
Not a general-purpose chatbot — asking unrelated questions produces garbage tool calls.
Small model, small context. Multi-turn conversations longer than a handful of exchanges will drift; lux avoids this by keeping history trimmed.
Destructive actions are not gated by the model — that's the job of the lux CLI, which asks for confirmation before running anything destructive. If you call the model directly, implement your own confirmation step.

License

Apache-2.0, inherited from Qwen3-1.7B. Fine-tuning derived weights are redistributed under the same terms.

Model tree for henrywangxf/lux

Base model

Qwen/Qwen3-1.7B-Base

Finetuned

Qwen/Qwen3-1.7B

Quantized

(258)

this model

henrywangxf
/

lux

lux