Recently, Meituan announced the launch of a large reasoning model based on the mixture of experts architecture - LongCat-Flash-Thinking. This new model has as many as 56 billion parameters, which is impressive! However, the most surprising thing is that it can dynamically activate between 18.6 billion and 31.3 billion parameters according to the context, with an average of about 27 billion parameters activated. This flexible design enables the model to perform well in various tasks, especially in areas such as logical reasoning, mathematical operations, and programming.
According to official data from Meituan, LongCat-Flash-Thinking has conducted in-depth comparisons with other mainstream models in a series of benchmark tests. In mathematical reasoning, for example, in the MATH500 and AIME25 tests, the model performed remarkably, even claiming that using native tools can reduce token consumption by 64.5% while still achieving top-level accuracy. Additionally, in areas such as general reasoning (GPQA-Diamond), code generation (LiveCodeBench, OJBench), and formal theorem proving (MiniF2F-Test), its performance is also close to industry-leading levels.
The model weights of LongCat-Flash-Thinking have been open-sourced under the MIT license, allowing developers to freely use and study this powerful tool. The official also provides detailed chat templates and has launched a dedicated chat website, making it convenient for users to interact and practice.
LongCat-Flash-Thinking is not just a large-scale AI model, but also an important attempt by Meituan in the field of artificial intelligence research. Its outstanding performance and flexible architecture will provide new possibilities for subsequent application development and research.
Official website: https://longcat.chat/
