智通财经网·2025-09-22 06:40Core Viewpoint - Meituan has launched a new efficient reasoning model called LongCat-Flash-Thinking, which demonstrates enhanced capabilities in tool invocation while maintaining a 90% accuracy rate and saving 64.5% of tokens compared to models that do not utilize tool invocation [1] Group 1 - The LongCat-Flash-Thinking model is based on AIME25 empirical data and has been fully open-sourced on platforms like HuggingFace and GitHub [1] - This model enhances the ability of intelligent agents to autonomously invoke tools and expands formal theorem proving capabilities, making it the first domestic large language model to combine "deep thinking + tool invocation" with "informal + formal" reasoning capabilities [1] - LongCat-Flash-Thinking shows significant advantages in handling highly complex tasks such as mathematics, coding, and intelligent agent tasks [1] Group 2 - Comprehensive evaluations indicate that LongCat-Flash-Thinking has reached the state-of-the-art (SOTA) level in reasoning tasks across various fields including logic, mathematics, coding, and intelligent agents [1]