我在 MacBook Air M5 上對 21 個本地大語言模型進行了程式碼品質和速度的基準測試

There are plenty of "bro trust me, this model is better for coding" discussions out there. I wanted to replace the vibes with actual data: which model writes correct code and how fast does it run on real hardware, tested under identical conditions so the results are directly comparable. No cherry-picked prompts, no subjective impressions, just pass@1 on 164 coding problems with an expanded test suite. Hardware: MacBook Air M5, 32 GB unified memory Quantization: Q4_K_M for all models via llam