In speculative decoding, the draft model is the cheap guesser: a model typically 10-100x smaller than the target that autoregressively proposes a handful of tokens, which the target then verifies in a single forward pass. It must share the target’s tokenizer, and the closer its training data and recipe are to the target’s, the more of its guesses survive verification.
The draft need not be a separate model at all: extra prediction heads on the target (Medusa, EAGLE, multi-token prediction) or even n-gram matches against the prompt can play the same role. What matters is only that guesses are cheap relative to the target’s forward pass, since wrong guesses are discarded by a rule that leaves the output distribution unchanged.