On July 6, Meituan unveiled its latest innovation—the open-sourcing of LongCat-2.0, a next-generation, trillion-parameter large model. This model boasts an impressive 1.6 trillion parameters, with an average activation of around 48 billion. LongCat-2.0 is tailored specifically for Agentic Coding tasks and comes with inference code that has been meticulously optimized for domestic computing power chips. On the same day, Moore Threads, a domestic GPU manufacturer, announced that it had achieved full-chain adaptation of LongCat-2.0 using its self-developed AI training and inference all-in-one GPU smart computing card, the MTT S5000, along with the MUSA software stack. This adaptation encompasses various aspects, including model loading, initialization of the inference engine, and optimization of key operators, ensuring stable and efficient inference of the model on the MTT S5000. Furthermore, several other domestic chip manufacturers, such as Huawei Ascend, have also accomplished similar adaptations.