Qwen3.6 35B-A3B on a 64 GB Mac: speed, quality, memory
Alibaba Qwen, released Apr 2026. 35B parameters, 3B active (mixture of experts). 4-bit MLX build, 20.4 GB download. License Apache-2.0.
- HumanEval+ pass@1: 152 of 164 (92.7%), rank 2 of 8
- HumanEval pass@1: 158 of 164 (96.3%)
- Generation speed: 86.5 tokens per second (5 trials, 84.0 to 88.0); prompt 1,373 tokens per second
- Peak memory: 20.7 GB · power while generating: 26 W · GPU peak 73 °C
Run it yourself
pip install mlx-lm mlx_lm.chat --model mlx-community/Qwen3.6-35B-A3B-4bit
HumanEval/99
Round a number string to the closest integer. Halfway cases round away from zero.
It used int(), which cuts toward zero. For -15.5 it compared -15 and -14, and returned -15 instead of -16.
Problems it failed
HumanEval/93, HumanEval/95, HumanEval/99, HumanEval/116, HumanEval/140, HumanEval/145
Raw files: speed.log, evaluate.log, eval_results.json
▶ Qwen3.6 35B-A3B on a 64GB Mac: 86 tok/s, 92.7% HumanEval+ (real test)