Meituan open-sources trillion-parameter large model LongCat-2.0, solving the domestic computing power training challenge, reducing failure rate by 70%. This wave of technical strength is worth paying attention to.

CoinWorld news, today, Meituan officially released the new generation trillion-parameter large model LongCat-2.0, which will be open-sourced. LongCat-2.0's pre-training data scale exceeds 30t tokens, covering Chinese, English, multilingual, code, and other types of data. Facing hardware failures, communication anomalies, memory pressure, and numerical fluctuations in ten-thousand-card-level training, the LongCat team has overcome the challenges of domestic computing power training from three aspects: stability, correctness, and efficiency. In terms of stability, through hccl exception handling, elastic scaling of cards, and automatic fault recovery, the monthly average daily failure rate is reduced by more than 70%. In terms of correctness, through self-developed deterministic operators, bitwise consistency verification, and parameter detection, the reliability of training results is ensured, while the calculation accuracy of key modules is improved and the reduce logic is optimized based on practice.