Drafter (Draft Model)

In speculative decoding, the drafter is a small, fast model whose job is to guess the next few tokens. It does not need to be perfect – it just needs to be quick and often right. The big model then verifies its proposals.

How many tokens the drafter proposes per step varies: the base model does 1, MTP does 2-3, and D-flash can go up to 7.