On April 24, Moore Threads, in partnership with the Zhiyuan FlagOS community, pioneered the Day-0 swift integration of the cutting-edge, large-scale model DeepSeek-V4-Flash on its premier AI training and inference all-in-one GPU, the MTT S5000. The company also carried out thorough optimizations and deployed all core operators. DeepSeek-V4-Flash utilizes a Mixture of Experts (MoE) architecture, boasting a staggering 284 billion parameters and accommodating a context window of up to one million tokens. As the inaugural full-function GPU in China to natively support FP8 precision, the Moore Threads MTT S5000 alleviates memory bandwidth strain by 50% and doubles the theoretical computing throughput, thanks to its hardware-level FP8 Tensor Core acceleration units. Presently, both entities are making strides in migrating and adapting DeepSeek-V4-Pro (with 1.6 trillion parameters) to the MTT S5000.