Day-0 support: Moore thread completes Meituan LongCat-2.0 ultra-fast adaptation

IT House reported on July 6 that today, Meituan officially open sourced LongCat-2.0, a large model with trillions of parameters. Moore Thread subsequently announced that it has completed rapid adaptation of the model based on the MTT S5000, an all-in-one full-featured GPU smart computing card for AI training and push, and the MUSA software stack. ▲ The LongCat-2.0 model runs on MTT S5000. According to reports, this adaptation covers the entire link including model loading, inference engine startup, key operator optimization, deployment verification and accuracy verification, enabling LongCat-2.0 to achieve stable and efficient inference operation on MTT S5000. As a new generation of trillion-parameter MoE large model self-developed by Meituan, LongCat-2.0 has a total parameter volume of 1.6T (average activation of about 48B, dynamic range of 33B~56B). This model is specially designed for Agentic Coding scenarios, natively supports 1M ultra-long context, and achieves efficient resource utilization and multi-task collaboration through self-developed sparse attention mechanism (LSA), ScMoE cross-layer fast connection architecture and zero-computing expert dynamic activation mechanism. Moore's thread technology team relies on the SGLang-MUSA inference engine and MUSA software ecosystem to quickly complete full-link adaptation from framework compatibility to performance optimization around the model structure and inference features of LongCat-2.0. Hardware-native FP8 support releases long-context reasoning performance. LongCat-2.0 is oriented to more complex task processing requirements. Actual deployment often involves reasoning loads such as long input, multi-round context, complex instruction decomposition and continuous generation, which puts forward higher requirements for GPU computing power, video memory capacity, memory access bandwidth and reasoning scheduling capabilities. MTT S5000 has hardware-level native FP8 acceleration capabilities. A single card has high computing power, large-capacity video memory, and high bandwidth, which can provide stable support for long context input, KV Cache reading and writing, and high-concurrency reasoning. Combined with SGLan