llama.cpp native MTP brings 1.4x–2.2x speedups on dense models, little on MoE

UsedMorning9886 · reddit · 2026-07-25

Native MTP in llama.cpp shows solid gains on dense models, weak ones on MoE

The post consolidates the current state of speculative decoding in llama.cpp after the dust settled.

Native --spec-type draft-mtp now lets models use their own built-in MTP heads instead of a separate draft model.
Qwen3.6, DeepSeek, and GLM ship with MTP heads that can use this path.
A prior speculative checkpointing merge in April laid the groundwork for reliable rollback on hybrid/recurrent architectures.
Reported results are mixed by architecture:
Dense models: meaningful speedups, often in the 1.4x–2.2x range for cases like Qwen3.6-27B dense.
MoE models: much smaller gains, sometimes none, because active-parameter cost per token is already low.
The author also notes that older speculative methods — separate draft models, n-gram matching, and similar tricks — have had inconsistent third-party verification, with some detailed benchmarks on Qwen3.6-35B-A3B showing no net speedup and even regressions.

The takeaway: if you want faster local inference today, native MTP heads look more reliable than older draft-model speculation.

More from coding & agent