lux-portable

Portable, single-download bundle of lux — an AI agent for the Linux desktop that manages your system through natural language.

This repo ships the standalone tarball: lux CLI + luxd daemon + llama-server (llama.cpp) + pre-downloaded fine-tuned weights. Extract and run — no system install, no network required at runtime.

Download

wget https://huggingface.co/henrywangxf/lux-portable/resolve/main/lux-portable-linux-x86_64-latest.tar.gz
tar xzf lux-portable-linux-x86_64-latest.tar.gz
cd lux-portable-linux-x86_64/
./lux

What's inside

lux             # interactive CLI / REPL
luxd            # background monitoring daemon (optional)
llama-server    # local inference server (llama.cpp)
libggml*.so     # shared libs loaded by llama-server ($ORIGIN rpath)
libllama*.so
libmtmd*.so
models/         # pre-downloaded Q4_K_M GGUF weights
systemd/        # reference unit file for running luxd via systemd --user

On startup, lux detects the sibling llama-server binary and models/*.gguf and auto-spawns a local inference server on an ephemeral port (with the model's chat template enabled via --jinja). The server terminates when lux exits. Nothing is written outside this directory.

Example session

lux> install firefox
Installing org.mozilla.firefox from Flathub...

lux> restart nginx
Restarting nginx.service...

lux> show me recent errors in the system log
Apr 17 01:22:15 fedora sshd[1234]: error: Authentication failure...

lux> my disk is almost full
Filesystem      Size  Used Avail Use% Mounted on
/dev/sda1       100G   89G   11G  89% /

Model

• Base: Qwen/Qwen3-1.7B (Apache 2.0)
• Fine-tuning: LoRA on a curated set of Linux sysadmin tool-use traces
• Format: GGUF Q4_K_M (served via llama-server)
• Intended use: parsing ambiguous natural-language requests into structured tool calls for lux's tool registry (install packages, manage services, read logs, manage firewall, etc.)

Most common queries hit lux's rule-based fast path and never reach the model. The model handles the long tail — ambiguous, novel, or multi-step requests.

Limitations

• English only.
• Tuned for Fedora / RHEL-family systems (dnf, systemd, firewalld). Other distros work but with reduced accuracy on package-manager queries.
• Not a general-purpose chatbot — trained exclusively for Linux system administration tasks.
• Running destructive commands (rm, dd, etc.) still requires your confirmation in the CLI.

License

• Model weights: Apache 2.0 (inherited from Qwen3).
• lux / luxd: Apache-2.0.
• llama-server + libggml*/libllama*/libmtmd*: MIT (bundled unmodified from llama.cpp).

Links

• Source: https://github.com/henrywang/lux
• Issues / feedback: https://github.com/henrywang/lux/issues
• Base model: https://huggingface.co/Qwen/Qwen3-1.7B