Meituan has unveiled LongCat, an open-source large language model that leverages a groundbreaking Mixture of Experts (MoE) architecture. Its flagship version, LongCat-Flash, impresses with 56 billion parameters and a swift reasoning speed of over 100 tokens per second. Through a meticulously designed multi-stage training pipeline and advanced optimization techniques, LongCat stands out in terms of training stability, scalability, and computational efficiency. In numerous benchmark tests, LongCat's performance rivals that of industry-leading models, particularly in areas such as the Massive Multitask Language Understanding (MMLU) benchmark, mathematical reasoning, and general reasoning tasks. By open-sourcing LongCat, Meituan aims to propel the advancement of AI technology and arm developers with a formidable tool to innovate.