据"华为计算"微信公众号29日消息，昇腾已快速基于vLLM/SGLang等推理框架完成适配部署，实现 DeepSeek-V3.2-Exp 0day支持，并面向开发者开源所有推理代码和算子实现。 寒武纪29日表示，已同步实现对深度求索公司最新模型DeepSeek-V3.2-Exp的适配。依托DeepSeek Sparse Attention机制，叠加寒武纪的极致计算效率，可大幅降低长序列场景下的训推成本。 9月29日，DeepSeek-V3.2-Exp模型正式发布并开源。模型引入稀疏Attention架构，据称能够有效降低计 算资源消耗并提升模型推理效率。此外，DeepSeek还大幅下调了API价格，降价幅度超过50%。 随后，华为昇腾、寒武纪、海光信息同步宣布完成适配。 30日上午，海光信息在官方微信公众号宣布，其DCU实现对DeepSeek-V3.2-Exp的无缝适配+深度调优， DeepSeek-V3.2-Exp在海光DCU上展现出优异的性能，同时验证海光DCU高通用性、高生态兼容度及自 主可控的技术优势。 受此消息提振，9月30日，市场早盘震荡拉升，科创50指数表现强势，其中AI芯片、华为昇腾概念表现 ...

转自：新华财经 新华财经北京9月30日电 9月30日早盘，AI相关概念股集体拉升。AI语料方向，当虹科技20CM涨停，开 普云、拓尔思、值得买等涨超5%；半导体硬件方向，德明利涨停，江波龙、联芸科技等涨超5%，寒武 纪、东芯股份等跟涨。 华鑫证券研报表示，国产AI芯片大时代已经来临，国产AI产业链从上游先进制程到先进封装，到下游 字节阿里腾讯的模型加速迭代升级已经实现全产业链打通，坚定看好国产AI算力设施的加速突破。 9月29日，深度求索公司正式发布新一代大语言模型DeepSeek-V3.2-Exp，并于HuggingFace平台同步开 源，模型引入稀疏 Attention 架构，这种架构能够降低计算资源消耗并提升模型推理效率。 据"华为计算"微信公众号消息，昇腾已快速基于vLLM/SGLang等推理框架完成适配部署，实现 DeepSeek-V3.2-Exp 0day支持，并面向开发者开源所有推理代码和算子实现。 寒武纪在官方微信号宣布，已同步实现对深度求索公司最新模型DeepSeek-V3.2-Exp的适配。依托 DeepSeek Sparse Attention机制，叠加寒武纪的极致计算效率，可大幅降低长 ...

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]

寒武纪表示，公司一直高度重视大模型软件生态建设，支持以DeepSeek为代表的所有主流开源大模 型。借助于长期活跃的生态建设和技术积累，寒武纪得以快速实现对DeepSeek-V3.2-Exp这一全新实验 性模型架构的day 0适配和优化。此前，公司对DeepSeek系列模型进行深入的软硬件协同性能优化，达 成了业界领先的算力利用率水平。针对本次的DeepSeek-V3.2-Exp新模型架构，寒武纪通过Triton算子开 发实现快速适配，利用BangC融合算子开发实现极致性能优化，并基于计算与通信的并行策略，再次达 成业界领先的计算效率水平。 "华为计算"公众号发文称，昇腾已快速基于vLLM/SGLang等推理框架完成适配部署，实现DeepSeek- V3.2-Exp 0day支持，并面向开发者开源所有推理代码和算子实现。华为昇腾在DeepSeek-V3.2-Exp一发 布开源即实现DeepSeek-V3.2-Exp BF16模型部署，并在CANN平台上完成对应的优化适配，整体部署策 略沿用DeepSeek的大EP并行方案，针对稀疏DSA结构，叠加实现长序列亲和的CP并行策略，兼顾时延 和吞吐，在128K长序列 ...

探索初显成效。算力调度底座打通：工业和信息化部《算力互联互通行动计划》明确2026年建全标准体 系，目前已完成131家企业499个算力资源池标识，华为昇腾实现毫秒级响应，和林格尔数据中心具 备"朝推夜训"调节能力。 伴随算力需求快速增长，其高耗电、高散热、高排放等问题日益凸显，对能源电力带来挑战。国家先后 出台算力能源电力协同相关文件，近日国家发展改革委、国家能源局又发布《关于推进"人工智能+"能 源高质量发展的实施意见》，进一步强调算电协同应用。算、电、热、碳耦合紧密，四者协同优化可释 放更大效能。 算—电—热—碳协同的必要性 算力高速发展带来"三高"压力。当前算力进入高速增长阶段，人工智能推动智算需求爆发，算力结构呈 现"智算引领、通算基础、超算补充"特点。我国算力近5年平均增速近30%，2024年智算增速高达 74%，远超通算的21%；IDC等机构预测，2023~2028年我国通算、智算规模复合增速为18.8%、 46.2%。依托超大型数据中心集群与边缘计算节点，我国算力通道持续为多领域提供支撑，但"三高"问 题突出。 高耗电：能源成为算力核心"生产要素"，剔除折旧后电费占数据中心运营成本60%以上；高 ...

Core Viewpoint - The company, Silver Trust Technology, confirmed its collaboration with Huawei's Ascend supernode by achieving certification for its self-developed intelligent operation and maintenance knowledge base version 1.2.0, and it is recognized as a certified application software partner in Huawei's Ascend Wanli Partner Program [1] Group 1 - The company has completed the development and certification of its intelligent operation and maintenance knowledge base version 1.2.0 [1] - The company is a certified application software partner in Huawei's Ascend Wanli Partner Program [1]

Group 1 - The core viewpoint of the article highlights the significant rise of the Sci-Tech 50 ETF, which increased by 4.05% with a trading volume exceeding 4 billion yuan, driven by strong performances from stocks like Chip Origin Technology, Haiguang Information, and Jingchen Technology, all rising over 10% [1] - Huawei announced its future development roadmap for Ascend at the Huawei Connect 2025 event, planning to launch several new products including the 950PR, 950DT, 960, and 970 between Q1 2026 and Q4 2028, along with the development of large-scale supernodes like Atlas 950 SuperCluster and Atlas 960 SuperCluster [1] - CITIC Securities expressed optimism about Huawei's Ascend as a representative of domestic computing power, recommending attention to advanced processes and the HBM and other supporting industrial chains [1] Group 2 - The Sci-Tech 50 ETF tracks the Sci-Tech 50 Index, which has a significant allocation in the electronics sector at 68.77% and the pharmaceutical and biological sector at 9.85%, totaling 78.62%, aligning well with the development directions of domestic chips, artificial intelligence, and robotics [1] - The index also covers multiple sub-sectors including semiconductors, medical devices, software development, and photovoltaic equipment, indicating a high content of hard technology [1] - The current position of the Sci-Tech 50 Index is near its base point, and referencing the historical performance of the ChiNext board, the future growth potential is promising, suggesting that investors optimistic about China's hard technology should continue to pay attention [1]

Core Insights - The report highlights the increasing risk of a government shutdown in the U.S. as the Senate Democrats rejected a temporary funding bill proposed by Republicans [4] - The report notes that OpenAI has reportedly reached an agreement with Luxshare Precision to produce future OpenAI devices, indicating a strategic partnership in technology development [4] - Huawei plans to invest 15 billion annually over the next five years to develop its ecosystem, including HarmonyOS, Kunpeng, and Ascend technologies, showcasing its commitment to innovation and market expansion [4] Economic Data - The Baltic Dry Index closed at 2203.00, down 0.09%, while the Nasdaq Index rose by 0.72% to 22631.48 [6] - The 2-year U.S. Treasury yield increased by 1.22 basis points to 3.567%, the 5-year yield rose by 1.92 basis points to 3.677%, and the 10-year yield increased by 2.49 basis points to 4.125%, reflecting market adjustments [4] - The Hang Seng Index remained unchanged at 26545.10, while the Shanghai Composite Index decreased by 0.30% to 3820.09, indicating mixed performance in the Asian markets [6]

Core Viewpoint - Hengwei Technology is planning to acquire a controlling stake in Shanghai Shuhang Information Technology Co., Ltd. to extend its AI strategy, marking the first case of an A-share listed company acquiring an AI RaaS target, which is seen as a strategic positioning move in the AI industry [1][5]. AI Strategy Extension - Hengwei Technology announced on September 16 that it is planning to purchase 75% of Shanghai Shuhang's shares through a combination of issuing shares and cash, while also raising matching funds through share issuance. This transaction is not expected to constitute a major asset restructuring [2]. - Shanghai Shuhang, established in 2017, began training a locally deployable S-GPT large model in 2022 and completed its localized deployment in March 2023. The company has developed a series of advanced AI applications based on customer demand scenarios [2]. - The acquisition aligns with Hengwei Technology's focus on upgrading and expanding its AI computing capabilities, as both companies have complementary strengths in AI infrastructure and application [2]. Innovative Business Model - The AI industry is still in its early stages, with few AI software and service companies achieving profitability. However, Shanghai Shuhang has already realized significant profitability, with rapid growth in revenue and profit over the past three years [3]. - Shanghai Shuhang is innovating its business model by practicing AI RaaS, which is seen as a potential final model for AI application implementation. This model shifts the value perception for customers from "buying tools" to "buying results," allowing for a significant increase in the revenue share from 1%-2% to over 20% [3]. Comprehensive Capability - Shanghai Shuhang has built a comprehensive capability from foundational models to practical applications, serving numerous brand companies across various industries such as fast-moving consumer goods, automotive, finance, education, and human resources [4]. - The company expects to see the effects of its data-driven business model starting in 2025, leading to increased revenue and improved profit margins through deep customer engagement and operational efficiency [4]. Strategic Positioning - The acquisition is not only about business synergy but also about capturing industry trends, with Hengwei Technology aiming to establish a significant first-mover advantage in the AI application sector [5][6]. - First movers in the AI application field benefit from advantages in technological advancement and customer loyalty, with the potential to create sustainable competitive barriers through deep integration of AI RaaS into client production systems [6].

Group 1 - The U.S. stock market continues to reach new highs, with Oracle rising over 4% while the Chinese concept index fell by 0.25% [1] - Southbound funds have seen a net inflow for 18 consecutive weeks, with 11 stocks experiencing a doubling in shareholding [2] - Several express delivery companies have raised collection prices in Shanghai, primarily targeting low-priced e-commerce items, with limited impact on personal shipments [1] Group 2 - Goldman Sachs indicates that the current rise in the Chinese stock market is based on healthier fundamentals and is sustainable [1] - Haichang New Materials plans to acquire 51% of Xinyi Communications for 255 million yuan [2] - ST New Power is under investigation by the China Securities Regulatory Commission for suspected information disclosure violations [2]