Meituan Unveils LongCat-2.0 Trillion-Parameter Model
Meituan has launched its new trillion-parameter model, LongCat-2.0, which will be open-sourced, according to reports on June 30. The model's pre-training data exceeds 30 trillion tokens, encompassing Chinese, English, multiple languages, and code. The LongCat team has addressed challenges in domestic computing power training, such as hardware failures and communication anomalies, by enhancing stability, accuracy, and efficiency. They achieved a 70% reduction in monthly failure rates through HCCL exception handling and automatic fault recovery. Additionally, they ensured training accuracy with deterministic operators and parameter checks, while optimizing key module precision and Reduce logic.
