Meituan has released its next-generation trillion-parameter large language model, LongCat-2.0, and said it will be open-sourced.
According to Jin10, the model’s pretraining dataset exceeded 30T tokens and covered Chinese, English, multiple languages, and code.
Meituan said the LongCat team addressed challenges in large-scale training using domestic computing power, including hardware failures, communication issues, memory pressure, and numerical fluctuations. The company said it focused on stability, correctness, and efficiency.
On stability, Meituan said it used HCCL exception handling, elastic scaling of cards, and automatic fault recovery, cutting the average monthly daily failure rate by more than 70%.
On correctness, Meituan said it developed deterministic operators, conducted bitwise consistency verification and parameter checks, and improved computational precision in key modules while optimizing Reduce logic to support reliable training results.