Core Breakthrough: From “Usable” to “Proficient” Tool Invocation
In current AI applications, the ability of models to invoke external tools (such as search engines, APIs, and databases) is key to completing complex real-world tasks. LongCat-Flash-Thinking-2601 aims to solve three core challenges in this process: how to accurately select among multiple tools, how to adapt to new tools cost-effectively, and how to handle errors when a tool call fails or data is missing. Through its optimized tool invocation mechanism, the model can execute tasks more autonomously and reliably, significantly improving its practicality in production environments.
First Open-Sourced “Rethinking Mode”
Another major highlight of this release is the first open-sourcing of the “Rethinking Mode.” This mode aims to simulate the human process of deep deliberation when facing complex problems. It consists of two core stages: first, “parallel thinking,” which involves simultaneously analyzing the problem from multiple perspectives and paths to broaden the solution space; second, “summarization and integration,” which consolidates and weighs the conclusions from different thought paths to form a more comprehensive and reliable decision. This capability shows significant advantages in tasks requiring a high degree of logic, such as multi-turn conversations, long-chain reasoning, and multi-tool collaboration.
Performance Evaluation: Outstanding Results on Multiple Benchmarks
According to data released by Meituan, LongCat-Flash-Thinking-2601 performed exceptionally well on a series of industry-standard benchmarks. Specifically, it scored 82.8 on the Large Code Benchmark (LCB) for coding ability and achieved a perfect score of 100 on the notoriously difficult AIME-25 mathematical reasoning test. Furthermore, the model reached SOTA levels in agentic tool use and search capabilities. The Meituan team pointed out that these results were not achieved by “cramming” for the tests but are thanks to an automated task synthesis pipeline, proving the model’s strong generalization and inference capabilities.
Open Ecosystem: Full Release of Weights, Code, and Online Demo
To promote the popularization and innovation of AI technology, the Meituan LongCat team has adopted a thoroughly open-source strategy. Developers can directly access the model weights of LongCat-Flash-Thinking-2601 for deployment and fine-tuning via major platforms like GitHub, Hugging Face, and ModelScope. They can also access its inference code to understand the model’s working principles. Additionally, the official website https://longcat.ai provides an online demo portal. This move significantly lowers the barrier for developers to use top-tier AI models, providing a solid foundation for building innovative AI applications.
