Qwen3.6-27B MTP boosts Apple silicon speed

Hot take: this is not a universal 2x speedup story, it is a tuning story, and on an M5 Max the gains look solid only when the draft model is actually being accepted often enough. The initial config was likely too conservative for speculative decoding, so the first MTP result under-represented the model’s upside. Raising spec-draft-n-max and setting spec-draft-p-min improved throughput materially, which points to draft quality and acceptance being the bottleneck. The coding prompt produced about 95% acceptance, which is why the MTP variant pulled ahead much more clearly there. The 27B numbers are the most useful data point here: 17.44 tps non-MTP versus 27.70 tps MTP is a meaningful improvement for local inference on Apple silicon. The takeaway for other users is to benchmark by workload, not just by model name, because prose and coding prompts can behave very differently with MTP.