Speculative decoding uses a lightweight draft model to generate candidate tokens that are then verified in parallel by a large model; it is a strictly lossless acceleration method that typically delivers a 2–3× speedup.