Meituan released LongCat-2.0, a new-generation trillion-parameter model, and
will open-source it. Pretraining used over 30T tokens covering Chinese, English,
multilingual data and code. Facing faults, communication anomalies, memory
pressure and numerical instability during ten‑thousand‑card–scale training on
domestic compute, the LongCat team focused on stability, correctness and
efficiency. For stability they implemented HCCL exception handling, elastic card
scaling and automatic fault recovery, cutting average daily failure rates by
more than 70%. For correctness they developed deterministic operators, bitwise
consistency verification and parameter checks, and improved key-module compute
precision and Reduce logic.
