Qwen3-Coder 30B-A3B on a 64 GB Mac: speed, quality, memory
Alibaba Qwen, released Jul 2025. 30.5B parameters, 3.3B active (mixture of experts). 4-bit MLX build, 17.2 GB download. License Apache-2.0.
- HumanEval+ pass@1: 134 of 164 (81.7%), rank 7 of 8
- HumanEval pass@1: 143 of 164 (87.2%)
- Generation speed: 93.9 tokens per second (5 trials, 93.2 to 94.1); prompt 1,402 tokens per second
- Peak memory: 17.8 GB · power while generating: 29 W · GPU peak 77 °C
Run it yourself
pip install mlx-lm mlx_lm.chat --model mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit
HumanEval/1
Split a string of nested brackets into its separate balanced groups.
The logic was right, but it used List without importing it. The code crashed with a NameError before a single test ran.
Problems it failed
HumanEval/1, HumanEval/65, HumanEval/75, HumanEval/81, HumanEval/89, HumanEval/100, HumanEval/109, HumanEval/113, HumanEval/115, HumanEval/116, HumanEval/120, HumanEval/122, HumanEval/123, HumanEval/127, HumanEval/129, HumanEval/130, HumanEval/132, HumanEval/145, HumanEval/159, HumanEval/160, HumanEval/163
Raw files: speed.log, evaluate.log, eval_results.json