Moore Threads Completes Full-Stack Adaptation for DeepSeek-V4

DeepSeek-v4">Another Win for Domestic GPUs: Moore Threads Successfully Runs DeepSeek-V4

On May 1, domestic GPU maker Moore Threads officially announced the completion of full-stack engineering adaptation and operational verification for DeepSeek's latest flagship model, DeepSeek-V4. The adaptation was built on Moore Threads' flagship AI training-and-inference compute card MTT S5000, its proprietary MUSA software stack, and the SGLang open-source inference framework, achieving end-to-end integration from underlying hardware to upper-level inference frameworks. This progress carries significant implications for the development of China's domestic AI computing ecosystem.

MTT S5000: A Domestic Compute Foundation for Training and Inference

The MTT S5000 is Moore Threads' flagship AI compute card designed for large-scale AI training and inference scenarios, positioned as a unified training-and-inference solution. In this adaptation, the S5000 demonstrated its capacity to support ultra-large-scale MoE (Mixture of Experts) architecture models. As the latest iteration in the DeepSeek series, DeepSeek-V4 features further increases in model scale and architectural complexity, placing extremely demanding requirements on the compute performance, memory bandwidth, and communication capabilities of underlying chips. Moore Threads' successful completion of full operational verification on the S5000 indicates that its hardware capabilities have met the fundamental requirements for supporting cutting-edge large models.

Synergy Between the MUSA Software Stack and SGLang Framework

Notably, this adaptation was not merely a hardware-level demonstration but encompassed deep integration between the proprietary MUSA software stack and the open-source SGLang inference framework. MUSA is Moore Threads' independently developed GPU computing software platform, serving a role analogous to CUDA in NVIDIA's ecosystem and representing a core component in building the software ecosystem for domestic GPUs. SGLang, meanwhile, is a popular high-performance inference framework in the open-source community, widely used for efficient deployment of large language models.

Moore Threads' decision to base the adaptation on SGLang reflects its strategy of actively embracing the open-source ecosystem while also lowering the barrier for developer migration — teams with existing SGLang experience can more quickly migrate their workloads to the Moore Threads platform.

A Critical Step for the Domestic Computing Ecosystem

Against the backdrop of persistent global tensions in AI chip supply, adaptation partnerships between domestic GPU manufacturers and domestic large models are accelerating. Previously, platforms such as Huawei Ascend and Hygon DCU had already completed support for mainstream domestic large models. Moore Threads' successful run of DeepSeek-V4 further enriches the model support matrix for domestic computing power.

However, a measured perspective is warranted: a gap still exists between "completing operational verification" and "achieving production-grade deployment readiness." Subsequent performance in areas such as inference speed, throughput, long-term stability, and multi-card parallel efficiency will be the key metrics for evaluating the real-world competitiveness of Moore Threads' solution.

Outlook

As next-generation large models like DeepSeek-V4 continue to drive up compute demands, domestic GPU manufacturers face unprecedented opportunities and challenges. Moore Threads' completion of this full-stack adaptation sends a positive signal: China's domestic AI chips are steadily progressing from "functional" to "performant." Looking ahead, there remains substantial work for domestic compute companies like Moore Threads in performance optimization, ecosystem development, and scaled commercial deployment — but the direction is unmistakably clear.