Meituan open-sources trillion-parameter AI model LongCat-2.0, now works with Moore Threads GPU

Meituan has open-sourced its self-developed trillion-parameter AI model LongCat-2.0, with a total of 1.6T parameters and an average of about 48B activated. This model is designed for Agentic Coding, supports 1M long context, and uses a custom sparse attention mechanism for efficient resource use.

Moore Threads says it has quickly adapted LongCat-2.0 to work with its MTT S5000 GPU and MUSA software stack. The adaptation covers model loading, inference engine, operator optimization, and deployment verification, ensuring stable and efficient operation on domestic computing platforms.

The MTT S5000 features hardware-level FP8 acceleration, with high computing power and large memory for long-context inference. With SGLang-MUSA inference engine optimization, it improves online service response and system throughput. Moore Threads has standardized the model adaptation process, lowering the barrier for deploying advanced models on domestic platforms.