On January 16, 2026, the LongCat team at Meituan made a significant announcement by officially launching and open-sourcing its cutting-edge agent model, named LongCat-Flash-Thinking-2601. This model represents an enhanced iteration of the LongCat-Flash-Thinking, reaching the pinnacle of performance—SOTA (State-of-the-Art)—among open-source models in crucial evaluation metrics such as agent search capabilities, tool invocation proficiency, and reasoning in tool interactions.

Notably, when tackling random complex tasks that hinge on tool invocation, LongCat-Flash-Thinking-2601 outperforms Claude, offering a substantial reduction in the costs associated with adapting and training for new tools.

One of the standout features of LongCat-Flash-Thinking-2601 is its support for a 'reflective thinking' mode. This innovative functionality allows for the concurrent activation of eight 'brains', fostering parallel thinking processes. Such an approach ensures thorough and all-encompassing thought processes, leading to more reliable and informed decision-making.

For those eager to explore the capabilities of this groundbreaking model, it is readily available for free trial on the official LongCat website.