我在相同的飛行模擬提示詞上測試了9個本地模型，全部Q8量化，不同的量化提供者，使用MLX

I gave 9 local models the same flight combat sim prompt. The results broke a few of my assumptions about quant providers and parameter count. All 8-bit MLX, M3 Max 128GB, served via omlx, prompted through Claude Code. Same prompt every time — single-file HTML, three selectable planes (jet, prop, wildcard of the model's choice), dynamic enemies, tracers, damage, crash spiral on loss. Counted prompts-to-final and graded on "does it actually play." https://alextzk.github.io/flight-combat-llm-comp