Qwen3.5-35B-A3B tops 63 t/s on M2 Ultra

This is a strong showing for a 35B-class MoE model on consumer-ish hardware, especially if you care about interactive local use more than leaderboard bragging rights.

• –The active-parameter design lets a large model fit and run fast enough on 64GB unified memory without immediately collapsing into tiny quants.
• –Prefill stays high even at 16K context, which matters more for long prompts and codebases than the raw generate number.
• –Q8_K_XL looks like a sensible sweet spot here: enough fidelity to keep the model interesting, without the memory hit of heavier formats.
• –Treat it as a hardware/backend benchmark, not a universal model ranking; no task suite, prompts, or quality scoring were reported.