Compare local coding models on a 64 GB Mac
Real measurements on a MacBook Pro · Apple M5 Pro · 64 GB. Speed from the official mlx_lm.benchmark, quality from EvalPlus HumanEval+ (164 problems, 124,253 test cases per model). Nothing is estimated.
| Rank | Model | HumanEval+ | HumanEval | Speed | Peak memory | License |
|---|---|---|---|---|---|---|
| 1 | Qwen3.6 27B | 93.3% | 96.3% | 17.6 tok/s | 17.0 GB | Apache-2.0 |
| 2 | Qwen3.6 35B-A3B | 92.7% | 96.3% | 86.5 tok/s | 20.7 GB | Apache-2.0 |
| 3 | Gemma 4 26B-A4B | 90.9% | 95.7% | 80.7 tok/s | 15.1 GB | Apache-2.0 |
| 4 | gpt-oss-20b | 87.2% | 92.7% | 82.6 tok/s | 12.7 GB | Apache-2.0 |
| 5 | Gemma 4 E4B | 83.5% | 87.2% | 80.5 tok/s | 5.1 GB | Apache-2.0 |
| 6 | Devstral Small 2 | 82.3% | 88.4% | 18.1 tok/s | 15.1 GB | Apache-2.0 |
| 7 | Qwen3-Coder 30B-A3B | 81.7% | 87.2% | 93.9 tok/s | 17.8 GB | Apache-2.0 |
| 8 | Gemma 4 E2B | 77.4% | 83.5% | 130.5 tok/s | 3.5 GB | Apache-2.0 |