flight-sim

Benchmark of 9 quantized local LLMs running MLX on a flight-combat HTML prompt shows quant provider choice and model quirks matter more than parameter count or bit-width for usable code output.