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)

bash

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

bash

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.

henrywangxf/lux

lux