MTP and speculative decoding

Multi-token prediction, or MTP, uses draft heads shipped with the model to propose more than one token at a time. Classic speculative decoding uses a separate compatible draft model. Both accelerate the same part of , so llm.fit applies the best supported method rather than multiplying them together.

Formula

accelerated decode tps = base decode tps × accepted speed multiplier

Worked example: Qwen3.6-27B with MTP

The research artifact records Qwen3.6-27B on an RTX 3090 rising from 38 to 65 tps with MTP. That is a 1.71 times gain and sits inside the artifact's 1.7 to 2.0 MTP range.

65 ÷ 38 = 1.71×
Try MTP on current Qwen3.5 27B