[Meituan Releases Next-Generation Trillion-Parameter Model LongCat-2.0] According to Jinshi reports, Meituan officially released its next-generation trillion-parameter model, LongCat-2.0, today and announced it will be open-sourced. The pre-training dataset for LongCat-2.0 exceeds 30T tokens, covering Chinese, English, multiple languages, and code, among other types of data. To address challenges such as hardware failures, communication anomalies, memory pressure, and numerical fluctuations during trillion-card-level training, the LongCat team tackled domestic computing power training difficulties from the perspectives of stability, accuracy, and efficiency.