The article presents TwT, a novel approach designed to enhance the reasoning capabilities of large language models while simultaneously reducing computational costs during inference. TwT achieves this by employing habitual reasoning distillation guided by multiple teachers, which streamlines the problem-solving process. This method focuses on distilling habitual reasoning patterns, allowing models to think without relying heavily on token-by-token processing. The goal of TwT is to make inference more efficient without compromising the quality of reasoning. By integrating guidance from multiple teacher models, TwT leverages diverse expertise to improve the distilled reasoning habits. This approach addresses the challenge of balancing performance and computational efficiency in large language models. Overall, TwT represents a promising advancement in optimizing reasoning processes for AI systems.