LLM Dogfight 0.5.0

LLM Dogfight 0.5.0 — first public release

A UFO dogfight in your terminal, flown by two small language models running on your own machine.

Install

cargo install --git https://github.com/AlveeeRahman/llm-dogfight   # Rust toolchain: https://rustup.rs
pip install torch transformers        # NVIDIA GPU (CUDA build of torch)   — or —
pip install mlx-lm                    # Apple silicon
dogfight                              # first run downloads the default pair (2.2 GB)

Prebuilt binaries for Linux x86_64 and macOS Apple silicon are attached below as tarballs with SHA-256 sums; put dogfight on your PATH and run it.

What you get

• Two local models command two saucer teams over a sleeping city: attack, hunt, flee, abduct cows, battle cries, and lessons learned from every loss. Default pair Qwen3-0.6B vs SmolLM2-360M (2.2 GB); a catalogue of 23 ungated models up to 14B, chosen with --zorb / --krell; dogfight models measures your GPU or unified memory and names the strongest pair it can hold.
• Rules that make a match: 4 saucers on screen, 20 reinforcements per game, a game ends when a team has nothing left, the loser returns with an extra saucer, scores and games won kept per model pair, one bolt in flight per saucer.
• Guardrails: a bounded tactical doctrine enforced on every order (AutoSafe-style threat model, reflection and safe-action correction); abduction is always the model's own call.
• dogfight evolve: a genetic algorithm evolves each team's doctrine, separately and in parallel, with no extra model calls.
• Runs anywhere text runs: plain truecolor text, no graphics protocol; Linux and macOS; CUDA or MLX; 0.8 MB Rust binary with one dependency.
• dogfight cpu plays with tiny models on no GPU at all, and a clean dogfight remove leaves nothing behind.
• Quality gates: rustfmt, clippy with every lint as an error, cargo audit, overflow checks, locked builds, pty harness on bash/zsh/fish, sidecar tests, bandit and ruff; a weekly Hugging Face check that every catalogue model is still public.

Requirements

• Linux or macOS, a truecolor terminal (GNOME Terminal, Ptyxis, kitty, iTerm2, Ghostty, …)
• For the model battle: Python 3.10+ with torch + transformers on an NVIDIA GPU (8 GB is enough for the defaults and most pairs), or mlx-lm on Apple silicon
• No GPU: dogfight ufo plays with the built-in pilots

Full Changelog: https://github.com/AlveeeRahman/llm-dogfight/commits/v0.5.0

Full Changelog: https://github.com/AlveeeRahman/llm-dogfight/commits/v0.5.0