Today, the LongCat team of Meituan officially released and open-sourced the latest AI model - LongCat-Flash-Thinking-2601. As an upgraded version of the LongCat-Flash-Thinking series, this model has achieved the state-of-the-art (SOTA) level in core evaluation benchmarks such as intelligent agent search, tool calling, and reasoning.

QQ20260116-135826.png

The core advantage of LongCat-Flash-Thinking-2601 lies in its exceptional tool calling capability. This ability enables the model to perform well in complex tasks that rely on tools, significantly reducing the training cost for adapting to new tools in real scenarios. In addition, the model's "rethinking mode" is provided for online free experience in an open-source format for the first time. Users can try it on the website https://longcat.ai. In this mode, the model simulates the process of human deep thinking, dividing the thinking into two stages: parallel thinking and summarization, ensuring comprehensive thinking and reliable decision-making.

After rigorous evaluation, LongCat-Flash-Thinking-2601 has shown excellent performance in multiple indicators such as programming, mathematical reasoning, intelligent agent tool calling, and search capabilities. In terms of programming ability, the model scored 82.8 points in the LCB evaluation, ranking among the top models in the same category; in mathematical reasoning, it achieved a perfect score of 100 points in the AIME-25 evaluation, further consolidating its leading position in this field.

QQ20260116-135848.png

To evaluate the model's generalization ability, the LongCat team also proposed a new evaluation method. This method uses an automated task synthesis process, supporting users to randomly generate complex tasks based on keywords and evaluate the model's performance in such environments. Experiments show that LongCat-Flash-Thinking-2601 maintains a leading performance in multiple randomly generated tasks, proving its strong generalization ability.

In the training process, the LongCat team adopted a strategy of "environment expansion + multi-environment reinforcement learning," providing the model with diversified high-intensity training environments, significantly improving its adaptability in complex scenarios. In addition, the team injected noise into the training data to enhance the model's robustness, allowing it to efficiently complete tasks even when facing complex situations such as API call failures or missing data.

To lower the development threshold, the LongCat team of Meituan also opened up the model's weights, inference code, and online experience capability, encouraging developers to actively participate in this open-source project. Developers can obtain resources through platforms such as GitHub, Hugging Face, and ModelScope, and experience it online at https://longcat.ai.