On July 6, Meituan made a significant announcement regarding its groundbreaking trillion-parameter large model, LongCat-2.0, officially releasing it as open-source. This release includes not only the model weights and inference engines but also comprehensive core technical documentation. On the same day, leading domestic chip manufacturers, including Huawei Ascend, Moore Threads, and MetaX, jointly declared the successful completion of inference adaptation work for this model. LongCat-2.0 is built upon a sophisticated Mixture of Experts (MoE) architecture, boasting a total parameter scale of 1.6 trillion. It natively supports ultra-long contexts, extending up to 1M, setting a new benchmark in the industry. During its development, the peak training scale surpassed 50,000 domestically produced computing cards, and the pre-training data scale exceeded an impressive 30 trillion Tokens. Notably, LongCat-2.0 stands as the industry's first trillion-parameter large model to accomplish the entire training and inference process exclusively utilizing domestic computing power, marking a significant milestone in technological self-reliance.