2025年，摩尔线程专注全功能GPU研发创新，持续推进产品架构迭代，成功推出旗舰级训推一体全功能GPU智算卡 MTTS5000，产品性能达市场领先水平并实现规模量产。基于该产品搭建的大规模集群已上线服务，可高效支持万亿参数大模 型训练，计算效率达到同等规模国外同代系GPU集群先进水平。 本报讯 （记者李乔宇） 2月27日晚间，摩尔线程智能科技（北京）股份有限公司（以下简称"摩尔线程"）披露2025年度业绩快报。2025年，摩尔 线程实现营业收入15.05亿元，同比增长243.37%；实现归属于母公司所有者的净利润为-10.24亿元，与上年同期相比亏损收 窄。此外，摩尔线程基本每股收益、加权平均净资产收益率同比均有所改善，整体展现出稳健向好的发展态势。 （编辑 张伟） 2026年春节前后，凭借MUSA架构卓越的生态兼容性和广泛的算子库，摩尔线程S5000已高效完成对GLM-5、 MiniMaxM2.5、KimiK2.5、Qwen3.5等SOTA大模型的深度适配。 ...

Core Viewpoint - "Tai Chi Yuan Qi" has completed deep adaptation work for several mainstream domestic open-source large models, indicating significant progress in AI model development and compatibility within the industry [1] Group 1: Company Developments - "Tai Chi Yuan Qi" has successfully adapted multiple mainstream domestic large models, including Zhipu GLM-5.0 and Alibaba Qwen 3.5-397B-A17B [1] - The adaptations were conducted on the company's self-developed T100 acceleration card, showcasing its technological capabilities [1] Group 2: Industry Impact - The introduction of a tiered development toolchain within the SDAA software stack aims to meet diverse development needs, from beginner to advanced levels [1] - This toolchain is designed to help developers quickly build high-performance operators, facilitating seamless compatibility with mainstream AI ecosystems [1] - The initiative significantly reduces the technical barriers and costs associated with migrating from the CUDA ecosystem [1]

Core Viewpoint - Moores Threads announced the official release of the new generation large model GLM-5, which has been fully adapted and verified on the flagship AI training and inference integrated GPU MTT S5000 on Day-0 [1] Group 1: Product and Technology - The MUSA architecture provides extensive operator coverage and strong ecological compatibility, enabling Moores Threads to successfully complete the full model inference chain [1] - The MTT S5000's native FP8 acceleration capability has been deeply leveraged, significantly reducing memory usage while ensuring model accuracy, thus achieving high-performance inference for GLM-5 [1] - The rapid adaptation of GLM-5 on MTT S5000 demonstrates the maturity of the MUSA software stack and showcases the support capabilities of domestic full-function GPUs for the latest large models [1] Group 2: Developer Experience - The combination of GLM-5 and MTT S5000 offers developers an exceptional programming experience that can compete with top international models [1] - This collaboration excels in various scenarios such as function completion, vulnerability detection, and debugging, significantly enhancing logical planning capabilities to tackle complex long-range task challenges [1]

Core Insights - Moore Threads has completed the full-process adaptation and verification of the latest large model GLM-5 on its flagship AI training and inference GPU, the MTT S5000 [1] Group 1: Product Development - The MTT S5000 is designed specifically for large model training, inference, and high-performance computing [1] - The GPU is built on the fourth-generation MUSA architecture, known as "Pinghu" [1] - The MTT S5000 offers a maximum AI computing power of 1000 TFLOPS per card [1]

Core Viewpoint - The release of the new generation large model GLM-5 by Zhiyu marks a significant advancement in AI capabilities, showcasing the effective integration of the MTT S5000 GPU with the SGLang inference framework for high-performance model inference [1] Group 1 - Zhiyu officially launched the GLM-5 model on February 11, demonstrating its capabilities in AI [1] - The MTT S5000 GPU achieved full-process adaptation and verification on Day-0, indicating rapid deployment capabilities [1] - The MUSA architecture provides extensive operator coverage and strong ecosystem compatibility, facilitating the complete model inference pipeline [1] Group 2 - The MTT S5000 GPU significantly reduces memory usage while ensuring model accuracy through its native FP8 acceleration capabilities [1] - The quick adaptation of the MTT S5000 not only validates the maturity of the MUSA software stack but also highlights the support capabilities of domestic full-function GPUs for the latest large models [1]

Core Viewpoint - The release of DeepSeek-V3.2-Exp model on Hugging Face platform introduces a sparse Attention architecture that reduces computational resource consumption and enhances inference efficiency [1] Group 1: Model Deployment and Adaptation - Huawei's Ascend has quickly adapted and deployed the DeepSeek-V3.2-Exp model based on vLLM/SGLang inference frameworks, providing open-source inference code and operator implementations for developers [1] - Cambricon announced the adaptation of the latest DeepSeek-V3.2-Exp model and has open-sourced the vLLM-MLU inference engine source code, leveraging the new DeepSeek Sparse Attention mechanism to significantly reduce training and inference costs in long-sequence scenarios [1] - Haiguang Information announced seamless adaptation and deep optimization of its DCU, achieving "zero-wait" deployment for large model computing power, showcasing excellent performance of DeepSeek-V3.2-Exp on Haiguang DCU [1]

Core Viewpoint - The emergence of compatibility issues between deployed AI models and chip architectures is becoming a hidden ceiling that restricts the practical application of AI, which Fourth Paradigm aims to address with its new solutions [1][7]. Group 1: Product Launch - Fourth Paradigm officially launched the "ModelHub XC" platform, the "Xinchang Community," and the "Xinchang Model Adaptation Value-Added Service" to tackle industry pain points and bridge gaps between customers, computing power, and developers [3]. - The "ModelHub XC" features an innovative AI engine system, EngineX, specifically designed to adapt to domestic computing power, fundamentally addressing the long-standing compatibility and support issues of domestic AI models [7]. Group 2: Market Context - Many existing ModelHubs primarily optimize foreign models and software for their hardware (e.g., NVIDIA GPUs), leading to compatibility issues with domestic hardware (e.g., Cambricon), resulting in time-consuming and repetitive adaptation processes [8]. - The platform has already certified and adapted over a hundred models upon launch, with plans to increase this number to thousands within six months and to reach tens of thousands within a year [10]. Group 3: Services and Support - Fourth Paradigm introduced a value-added service for model adaptation, providing tailored adjustments for users unfamiliar with which models are compatible with domestic computing power, ensuring a "safety net" for model compatibility [12]. - The platform also offers clear labeling of compatible domestic chip brands for each model, simplifying the process for users to determine which chips to purchase based on the models they wish to download [10].