On January 16, 2026, the LongCat team at Meituan officially announced and open-sourced the latest iteration of their model, LongCat-Flash-Thinking-2601. Serving as an enhanced version of LongCat-Flash-Thinking, this model reaches state-of-the-art (SOTA) performance levels among open-source models in crucial assessments, including agent search, tool calling, and reasoning for tool interaction. Specifically, when it comes to the generalization ability in tool calling, its performance outstrips that of Claude when tackling random complex tasks. This advancement substantially cuts down on the adaptation and training expenses associated with integrating new tools in practical settings. The model incorporates a 're-thinking' feature, enabling it to simultaneously engage 8 'cognitive processes' to guarantee comprehensive analysis and dependable decision-making. You can access its online experience at https://longcat.ai.