LLM Dogfight is a terminal game in which two small language models, running on your own machine, each command a team of UFOs. Team ZORB takes orders from one model and team KRELL from another. Every second, each commander gives every saucer an order: attack, hunt, flee or abduct a cow. After a loss, the commanders write short lessons for themselves.
The game is a single Rust binary with one dependency (libc). The models run in a Python sidecar on an NVIDIA GPU with CUDA, on Apple silicon with MLX, or slowly on the CPU. With no GPU at all, dogfight ufo flies the built-in pilots instead.
How a battle works
Each team has four saucers on screen and 20 reinforcements per game. The losing team fields one extra saucer in the next game. A doctrine of six hard-bounded numbers, such as when a saucer must flee, keeps orders sane: an order that breaks the bounds is corrected before it is flown. Abduction decisions are never corrected.
dogfight evolve adds a genetic algorithm that evolves each team's doctrine from game to game. The models themselves are never fine-tuned.
Install
cargo install --git https://github.com/AlveeeRahman/llm-dogfight
pip install torch transformers # or: pip install mlx-lm
dogfight # the first run downloads the default models (2.2 GB)
CUDA needs an NVIDIA GPU with 8 GB or more. The default pair is qwen3-0.6b for ZORB and smollm2-360m for KRELL, and dogfight models recommends the strongest pair your hardware can run. Any Hugging Face model with a chat template and safetensors weights also works.
Safety
The binary makes no network connections itself. Models are downloaded by the Python sidecar through huggingface_hub, and they never get tools, files or a shell. Model output is reduced to printable ASCII before it reaches your terminal. dogfight remove deletes everything the tool installed.