Speculative decoding for local LLMs: how much do MTP and DFlash really speed things up?

Speculative decoding lets a cheap drafter propose several tokens that the big model then verifies in one pass. Qwen3.8 ships its own MTP head, and DFlash drafters exist for several models. RigBase lists 25 measured pairs from 18 builds where the same model was run on the same machine with and without it.

Summary by method

Speed-up = speculative result ÷ the fastest plain-decode result for the same model on the same build. Both numbers come from the same source write-up; prompts, quants and engines can differ between the two rows, so open the build page before quoting a single ratio.

Every measured pair, fastest speed-up first

How to read this

• Dense models benefit most. One verification pass reads the weights once and serves several tokens. MoE models route different tokens to different experts, so verification reads more weights.
• Prompt type matters. Code is predictable and accepts many drafted tokens; open prose accepts fewer. Several sources report code and prose separately, and RigBase lists the lower figure as the headline.
• Don't compare across modes. A speculative number is not a like-for-like answer to a plain-decode number on another rig. Filter the builds table by speculative-decoding tag, or use the VRAM tiers, which show plain results first.
• Context and power. Gains shrink at deep context and when GPUs are power-capped; the build pages list both where the source states them.

Related: Is Q4 good enough? · GGUF vs AWQ vs FP8 vs NVFP4 · Qwen3.8 27B results · Qwen3.8 Flash-Next results · free dataset

FAQ

How much faster is MTP speculative decoding on a home GPU?

Across 20 measured MTP pairs on RigBase (same build, same model, with and without MTP) the median speed-up is 1.70x, ranging from 1.16x to 3.51x. Dense models gain more (1.84x median over 12 pairs) than mixture-of-experts models (1.40x median over 8 pairs).

Can speculative decoding make a model slower?

Yes. 1 of the 25 pairs below is slower with speculative decoding than without it. It is most likely on mixture-of-experts models, where verifying several drafted tokens reads many more expert weights, and on prose where few drafted tokens are accepted.

Is the output the same with and without MTP or DFlash?

In theory yes, because the big model verifies every drafted token. In practice numerical differences can change near-tied tokens; one measured Mac test found different greedy text in 7 of 8 scenarios (see its build page) and did not judge quality. RigBase lists speed only.

Which speed should I compare my own setup to?

Compare like with like: plain decode against plain decode, MTP against MTP. RigBase tags every result with its speculative-decoding mode and the how-it-was-measured line on each build page.