Meituan LongCat team releases new open-source AI model, leading the new trend in the open-source community
Today, the LongCat team under Meituan officially announced the launch and open sourcing of their latest masterpiece - LongCat Flash Thinking 2601. As an upgraded version of the LongCat Flash Thinking series, this model has reached the top level of the current open source model (SOTA) on many core evaluation benchmarks such as agent search, tool invocation and reasoning ability, which undoubtedly injects a strong new power into the open source community.
Excellent tool calling capability, reducing the cost of adapting new tools
The core highlight of LongCat Flash Thinking 2601 lies in its excellent tool calling capability. When faced with complex tasks that rely on tools, the model demonstrates extraordinary adaptability and flexibility, being able to quickly understand and call appropriate tools to complete the task.
This feature not only significantly enhances the application value of the model in real-world scenarios, but also significantly reduces the cost of adapting and training new tools, providing developers with a more convenient and efficient development experience.
The 'Rethinking Mode' is open source for the first time, simulating the process of human deep thinking
It is worth mentioning that LongCat Flash Thinking 2601 also provides an online free experience of "rethinking mode" in open source form for the first time. Users only need to visit the Longcat website to personally experience the charm of this innovative model.
In this mode, the model simulates the process of human contemplation, dividing it into two stages: parallel thinking and summarization. This phased thinking approach ensures the comprehensiveness of thinking and the reliability of decision-making, providing users with more accurate and reliable reasoning results.
Excellent multiple evaluation indicators, outstanding programming mathematical reasoning ability
After rigorous evaluation and testing, LongCat Flash Thinking 2601 has demonstrated outstanding performance in programming, mathematical reasoning, agent tool invocation, and search capabilities. In terms of programming ability, the model achieved an excellent score of 82.8 in the LCB evaluation, ranking among the top models in its class; In terms of mathematical reasoning, the model achieved a perfect score of 100 in the AIME-25 evaluation, further consolidating its leading position in this field.
Innovative evaluation methods validate generalization ability, with leading performance in random tasks
In order to comprehensively evaluate the generalization ability of the model, the LongCat team has also proposed a new evaluation method. This method utilizes an automated task synthesis process, supports users to randomly generate complex tasks based on keywords, and evaluates the performance of the model in such environments. The experimental results show that LongCat Flash Thinking 2601 maintains a leading performance in multiple randomly generated tasks, fully demonstrating its strong generalization ability and adaptability.
Unique training strategies and data augmentation to enhance adaptability to complex scenarios
During the training process, the LongCat team adopted a unique strategy of "environment extension+multi environment reinforcement learning" to provide the model with diverse high-intensity training environments. This training method significantly improves the model's adaptability in complex scenarios, enabling it to better cope with various challenges.
In addition, the team also injected noise into the training data to enhance the robustness of the model. This measure enables the model to efficiently complete tasks even in complex situations such as API call failures or data loss, demonstrating its strong fault tolerance and robustness.
Open resources lower development barriers and encourage developers to actively participate
In order to lower the threshold for developers to use and promote the prosperous development of the open source community, the Meituan LongCat team has also opened up the model's weights, inference code, and online experience capabilities. Developers can easily access these resources through platforms such as GitHub, Hugging Face, and ModelScope, and experience them online on the longcat.ai website.
This measure undoubtedly provides developers with more convenient and efficient development tools and environments, encouraging them to actively participate in this open source project and jointly promote the development and progress of AI technology.
Conclusion: Open source sharing, building the future of AI together
The release and open source of Meituan LongCat Flash Thinking 2601 not only demonstrates Meituan's profound strength and innovative capabilities in the field of AI, but also brings new vitality and opportunities to the open source community.
We look forward to more developers joining this open source project to explore the infinite possibilities of AI technology and work together to build a better future for AI.
