Lates News

Today Meituan officially released the new generation trillion-parameter large-scale model LongCat-2.0 and will open source it to the public. The pre-training data scale of LongCat-2.0 exceeds 30 trillion tokens, covering Chinese, English, multiple languages, and code data; in the face of hardware failures, communication anomalies, memory pressure, and numerical fluctuations during training at the level of ten thousand cards, the LongCat team overcame the training difficulties of domestic computing power from the three aspects of stability, correctness, and efficiency. In terms of stability, through HCCL abnormal handling, elastic scaling cards, and automatic fault recovery, the monthly average fault rate was reduced by more than 70%; in terms of correctness, through self-developed deterministic operators, Bitwise consistency verification and parameter detection to ensure the reliability of the training results, while improving the calculation accuracy of key modules and optimizing Reduce logic based on practice.