海光DCU率先支持DeepSeek-V3.2-Exp

Core Viewpoint - DeepSeek-V3.2-Exp has been released and open-sourced, introducing a sparse Attention architecture, with Haiguang DCU achieving seamless adaptation and deep optimization for zero-wait deployment of large model computing [1] Group 1: Company Developments - Haiguang Information is committed to building an AI software stack ecosystem under the "Deep Computing Intelligence" strategy, fully supporting the global mainstream open-source large models led by DeepSeek [1] - Haiguang DCU has rapidly completed "Day0" level efficient adaptation and optimization for DeepSeek-V3.2-Exp, benefiting from long-term and active technological accumulation [1] Group 2: Technical Performance - DeepSeek-V3.2-Exp demonstrates excellent performance on Haiguang DCU, leveraging the strong ecological advantages of the GPGPU architecture and the characteristics of the programming development software stack DTK [1] - The successful performance of DeepSeek-V3.2-Exp on Haiguang DCU validates the high versatility, high ecological compatibility, and self-controllable technological advantages of Haiguang DCU, establishing it as a key infrastructure for AI large model training and inference [1]