Glossary · Architecture

Draft Model

In speculative decoding, the smaller, faster model that proposes candidate next tokens which a larger 'verifier' model then checks, speeding up generation without changing the final output quality.