【Tech 24H】On September 1, Meituan officially released and open-sourced LongCat-Flash-Chat. Utilizing an innovative Mixture-of-Experts (MoE) architecture, the model boasts a total of 560B parameters, with activated parameters ranging from 18.6B to 31.3B (averaging 27B), achieving dual optimization of computational efficiency and performance. According to comprehensive evaluations across multiple benchmarks, LongCat-Flash-Chat demonstrates outstanding advantages in agent tasks while activating only a small number of parameters, with inference speeds exceeding 100 tps.