Cursor employs 'Speculative Decoding' technology, using a small draft model to first generate candidate completions, which are then verified or corrected by a larger model.