mHC新架构
Search documents
AI 系列跟踪(88):AI 芯片厂商密集上市,DeepSeek 提出新架构,AI 产业化进程再加速
Changjiang Securities· 2026-01-06 11:10
Investment Rating - The report maintains a "Positive" investment rating for the industry [7] Core Insights - Recent developments in the AI sector include the successful listing of Wallen Technology on the Hong Kong Stock Exchange and Baidu's Kunlun Chip planning a spin-off listing. DeepSeek has proposed a new mHC architecture that reduces the energy and computational requirements for training advanced AI, potentially accelerating the industrialization of AI [2][4] - The report highlights the upcoming IPOs of AI companies Zhiyu and MiniMax on January 8 and 9, respectively, and notes the partnership of Doubao with the Spring Festival Gala as a significant event [2][10] - The report identifies several promising investment opportunities within the AI sector, including high-quality IP benefiting from AI technology advancements, internet giants with advantages in traffic, models, and data, and vertical sectors like advertising, e-commerce, and education that have successfully replicated overseas business models in China [2][10] Summary by Sections Recent Events - Wallen Technology has successfully listed on the Hong Kong Stock Exchange, filling an important gap in the computing power sector. The company has developed a full chain of capabilities from high-end AI chips to computing clusters, with its self-developed "Biren" GPGPU architecture and related hardware products. The stock surged by 75.82% on its first day, indicating a new phase for the domestic computing power industry [10] - Baidu's Kunlun Chip is set to enhance its valuation transparency and attract investors focused on hard technology by planning a spin-off listing. The Kunlun Chip P800 cluster, capable of supporting multiple large models, marks a significant milestone in domestic computing power [10] - DeepSeek's new mHC architecture addresses issues in the existing Hyper-Connections structure, showing a mere 6.7% increase in training time while achieving significant performance improvements, thus lowering the costs associated with AI model training [10] Investment Opportunities - The report emphasizes the accelerated marginal growth in AI, with a focus on investment opportunities in the AI sector. It highlights the potential of high-quality IP benefiting from AI advancements, internet giants with data advantages, and vertical sectors that can replicate successful overseas business models [2][10]
月之暗面计划今年初上线多模态新模型;智元发布一体化具身大小脑系统GenieReasoner丨AIGC日报
创业邦· 2026-01-02 01:09
扫码可订阅产业日报 1.【月之暗面计划今年初上线多模态新模型】1月1日消息,记者获悉,月之暗面计划今年一月或者三 月上线多模态新模型,型号或为K2.1/K2.5。( 科创板日报 ) 2.【 DeepSeek 元旦发布新论文 , 开启架构新篇章】 DeepSeek 在元旦发布了一篇新论文,提出 了一种名为 mHC (流形约束超连接)的新架构。该研究旨在解决传统超连接在大规模模型训练中的 不稳定性问题,同时保持其显著的性能增益 。这篇论文的第一作者有三位: Zhenda Xie (解振 达)、 Yixuan Wei (韦毅轩)、 Huanqi Cao 。值得注意的是, DeepSeek 创始人 &CEO 梁文 锋也在作者名单中。(凤凰网) 3.【智元发布一体化具身大小脑系统GenieReasoner】1月1日,智元具身研究中心宣布推出第二代一 体化具身大小脑系统GenieReasoner。针对VLA模型中语义推理与动作控制的模态对齐难题,智元具 身研究中心提出了一种支持统一离散化预训练的模型架构,并通过流匹配(Flow-matching)缓解了 传统离散Tokenizer的动作精度瓶颈。同时具身研究中心开源了ER ...