Qwen3.6-35B-A3B smokes Qwen3.5 locally

The takeaway is blunt: for local MLX users, the newer sparse 35B model looks much more practical than the much larger 397B class model, even before you factor in memory pressure and responsiveness.

• –In LM Studio, Qwen3.6-35B-A3B hits 70 t/s versus 25 t/s for Qwen3.5-397B-A17B, with time-to-first-token dropping from 43.41s to 3.8s
• –In oMLX, it still holds a wide lead: 55 t/s versus 21 t/s, and 3.74s TTFT versus 25.6s
• –Prompt processing is the real standout: roughly 2400 t/s on Qwen3.6 versus 210-356 t/s on the older model
• –This is a single-user benchmark, so it is not a controlled eval, but it strongly suggests the newer model is the better local default for long prompts
• –For Mac/MLX deployments, speed at prompt ingest and first token matters more than headline parameter count, and this result favors the smaller sparse architecture