Comparing Qwen on M5 Pro 64GB

On a MacBook Pro M5 Pro 18-core, 64GB, Qwen 3.6 35B A3B 4bit and 27B dense 6bit were compared using LM Studio with the MLX runtime. Settings were thinking OFF (/no_think) and 128K context.

• Model size: 35B A3B was about 21.7GB, 27B dense was about 30.5GB
• 128K memory usage: 35B A3B was about 27GB, 27B dense was about 38GB
• Speed: in an 800-token test, 35B A3B ran at about 72 tok/s, 27B dense at about 9 tok/s — a difference of about 8x
• In a 1200-token test as well, 35B A3B recorded about 70 tok/s, 27B dense about 9 tok/s

Across 4 coding benchmark tasks, 35B A3B was generally superior overall.

• Auth hook: 9.5/10 vs 8/10
• Conflict resolution: 10/10 vs 10/10
• Delete account: 10/10 vs 10/10
• Bug identification: 10/10 vs 7/10
• Total score: 9.8/10 vs 8.75/10

The author concluded that on a 64GB Apple Silicon setup, 35B A3B was superior in both speed and quality, and dense 27B did not show the reasoning advantage that was expected.