A small model drafts, a big model checks several tokens at once, and the output distribution stays exactly the same. How and when it speeds things up.
