Base, Instruct, Thinking variants. 131K context. Thinking scores 78.4% EvalPlus. 2x+ faster inference vs similar dense models.