Skip to content
Technology

Speculative Decoding

A speed optimization where a small, fast 'draft' model generates candidate tokens that a larger model then verifies in parallel. Achieves near-large-model quality at small-model speeds — 2-3x faster generation.

Why it matters

Large models are slow because they generate one token at a time. Speculative decoding speeds this up by letting a small, fast model guess several tokens ahead, which the big model then checks all at once. When the guesses are right, you skip a lot of waiting. It's a clever trick that makes powerful models feel snappier without lowering the quality of the final output.

A concrete example

Picture a fast intern drafting the next few words of a sentence and a careful editor reviewing them in one pass. If the editor agrees, those words are accepted instantly instead of being written slowly one by one; if not, the editor corrects them. That's speculative decoding: the small draft model proposes, the large model verifies, and correct guesses let responses come out two to three times faster.

How to use it

Mostly invisible unless you host models yourself, where it is one of the standard ways to cut latency without changing the output distribution — a small model proposes tokens and the large one verifies them in parallel. If you are choosing an inference server, whether it supports this is a reasonable thing to compare, since the speed-up on long generations is real. The technique is set out in Fast Inference from Transformers via Speculative Decoding.

The common mistake

Assuming faster means a different, worse answer. The technique is designed to produce the same output the large model would have produced; the gain is in how the tokens are computed, not in what they are.

Related terms

Put Speculative Decoding into practice

Access 750+ AI models and 60+ tools through Vincony — start free with 100 credits.