On July 6, Meituan announced the official open-sourcing of LongCat2.0, including model weights, inference engine, and core technical documentation. On the same day, domestic chip manufacturers including Huawei Ascend, Moore Threads, and MetaX stocks collectively announced that they had completed inference adaptation for the model.
It is reported that LongCat2.0 adopts an MoE (Mixture of Experts) architecture, with a total parameter scale of 1.6 trillion. Every token activates an average of about 48 billion parameters per token, native support for the 1M ultra-long context. The training peak scale exceeds 50,000 domestic computing cards, and a pre-training data scale exceeds 30 trillion tokens.