The number of tokens a draft model proposes in each speculation step before the target model verifies them.