Meituan has announced today via the official Weibo account of "LongCat" that it has released and open-sourced LongCat-Flash-Chat. LongCat-Flash-Chat utilizes an innovative mixture-of-experts (MoE) model architecture, boasting an impressive total of 560 billion parameters. The activated parameters range from 18.6 billion to 31.3 billion, averaging at 27 billion, resulting in a dual enhancement in both computational efficiency and performance. Evaluated through a series of rigorous benchmark tests, LongCat-Flash-Chat has demonstrated exceptional performance in agent tasks, requiring only a fraction of activated parameters to achieve an inference speed exceeding 100 transactions per second. This accomplishment underscores Meituan's profound technical expertise in the realm of artificial intelligence.