DeepSeek正在改写AI的扩展法则：下一代AI的决胜点不再是单纯堆砌更大的GPU集群，而是通过更聪明的混合架构，用性价比更高的DRAM置换 稀缺的HBM资源。 据追风交易台消息，摩根士丹利1月21日发布的最新研报显示，DeepSeek正在通过一种名为"Engram"的创新模块，改变大语言模型的构建方式。 其核心突破在于将存储与计算分离，通过引入"条件记忆"（Conditional Memory）机制，大幅减少了对昂贵且紧缺的高带宽内存（HBM）的需 求，转而利用成本更低的普通系统内存（DRAM）来处理复杂的推理任务。 DeepSeek的解决方案是引入"条件记忆"（Conditional Memory）原则，即Engram模块。 这一架构的核心在于将静态模式存储与动态推理分离。DeepSeek不再将所有信息一次性加载到昂贵的HBM中，而是将模型的"图书馆"或"字 典"（静态知识）卸载到CPU或系统内存（DRAM）中，仅在需要时进行检索。 大摩分析师在报告中强调："DeepSeek将'条件记忆'与计算分离，为大语言模型（LLM）解锁了新的效率水平。Engram是一种在不通过重载HBM 的情况下，高效'查找'基 ...

DeepSeek正在改写AI的扩展法则：下一代AI的决胜点不再是单纯堆砌更大的GPU集群，而是通过更聪明的混合架构，用性价比更高的DRAM置换 稀缺的HBM资源。 据追风交易台消息，摩根士丹利1月21日发布的最新研报显示，DeepSeek正在通过一种名为"Engram"的创新模块，改变大语言模型的构建方式。 其核心突破在于将存储与计算分离，通过引入"条件记忆"（Conditional Memory）机制，大幅减少了对昂贵且紧缺的高带宽内存（HBM）的需 求，转而利用成本更低的普通系统内存（DRAM）来处理复杂的推理任务。 大摩分析师Shawn Kim及其团队认为，DeepSeek展示了如何"少花钱多办事"（Doing More With Less）的哲学。这种将存储与计算分离的技术路 径，不仅缓解了中国面临的AI算力约束，更向市场证明了高效的混合架构才是AI的下一个前沿。 这一被大摩重点关注的架构，源自DeepSeek创始人梁文锋团队与北大合作者在1月13日发布的重磅论文《Conditional Memory via Scalable Lookup》。在这篇论文中，团队首次提出了"Engram"（印迹）模块。 ...

Summary of DeepSeek's Innovation and Investment Implications Company and Industry Overview - **Company**: DeepSeek, a China-based AI company - **Industry**: Artificial Intelligence (AI) and semiconductor technology Core Insights and Arguments 1. **Innovation in AI Architecture**: DeepSeek's Engram module reduces high-bandwidth memory (HBM) constraints and infrastructure costs by decoupling storage from compute, suggesting that future AI advancements may focus on efficient hybrid architectures rather than merely larger models [1][2][9] 2. **Efficiency Gains**: The Engram approach enhances efficiency for Large Language Models (LLMs) by allowing essential information retrieval without overloading HBM, potentially reducing the need for costly HBM upgrades [2][3] 3. **Performance Metrics**: DeepSeek's findings indicate that hybrid architectures can outperform traditional models, with a minimum requirement of around 200GB system DRAM compared to existing systems that utilize significantly more [3][12] 4. **Next Generation LLM**: The upcoming DeepSeek LLM V4 is expected to leverage the Engram architecture, particularly excelling in coding and reasoning tasks, and may run efficiently on consumer-grade hardware [4][5] Investment Implications 1. **Market Potential**: Despite China's AI market being smaller than that of the US, its growth momentum suggests that investment opportunities may be underestimated. The report favors investments in Chinese memory and semiconductor localization themes, highlighting companies like Naura, AMEC, and JCET [5][9] 2. **Strategic Positioning**: By focusing on algorithmic efficiency rather than hardware expansion, DeepSeek exemplifies how companies can navigate geopolitical and supply-chain constraints, potentially leading to a more cost-effective and scalable AI ecosystem in China [21][16] Additional Important Insights 1. **Performance Comparison**: Over the past two years, Chinese AI models have significantly closed the performance gap with leading models like ChatGPT 5.2, emphasizing efficiency-driven innovations rather than sheer parameter growth [10][16] 2. **Conditional Memory Concept**: Engram introduces a method to separate static memory from dynamic reasoning, optimizing GPU usage and enhancing long-context handling, which has been a challenge for many large models [11][24] 3. **Benchmark Performance**: Engram has shown improved performance in benchmark tests, particularly in handling long-context inputs, which enhances the utility of AI models [20][21] This summary encapsulates the key points from the conference call regarding DeepSeek's innovations, their implications for the AI industry, and potential investment opportunities in the context of China's evolving AI landscape.

Group 1 - The core viewpoint of the news highlights the significant growth and penetration of artificial intelligence (AI) in various industries, with projections indicating that the number of AI companies in China will exceed 6,000 by 2025 and the core industry scale is expected to surpass 1.2 trillion yuan [1][2] - As of January 20, 2026, AI has penetrated over 70% of business scenarios in leading smart factories, with more than 6,000 vertical models developed, driving the large-scale application of over 1,700 key intelligent manufacturing equipment and industrial software [1] - The AI applications have covered key industries such as steel, non-ferrous metals, electricity, and telecommunications, gradually deepening into critical areas like product development, quality inspection, and customer service [1] Group 2 - The DeepSeek-R1 model has seen the emergence of a new model named "MODEL1" in the open-source community, indicating ongoing advancements in AI technology [2] - Industry experts predict that the global large model sector will continue to accelerate, with strong competitive advantages for China's AI development, as major tech companies are expected to enhance their capital expenditures to support model upgrades [2] - The Southern China AI ETF closely tracks the performance of the AI index, which reflects the stock price changes of listed companies related to the AI theme, with the top ten weighted stocks including companies like Zhongji Xuchuang and Tianfu Communication [2]

Group 1: Technology Developments - DeepSeek has updated its GitHub repository, revealing a new model architecture "MODEL1," which is expected to be more efficient and suitable for edge devices compared to its predecessor DeepSeek-V3.2 [2] - Longji Technology announced significant progress in Co-packaged Optics (CPO) technology, with successful customer sample deliveries and testing, addressing the growing demand for high-bandwidth, low-latency optical interconnects [11] - Shanghai Yiyou Intelligent Control Technology has launched its first automated production line for robot joints in Zhangjiang, aiming to meet the increasing demand and reduce costs for humanoid robots [10] Group 2: Financial Performance and Projections - Moole Technology expects a net loss of 950 million to 1.06 billion yuan for 2025, despite launching a leading GPU product and experiencing revenue growth due to the AI industry's expansion [17] - Demingli anticipates a net profit of 650 million to 800 million yuan for 2025, representing a year-on-year increase of 85.42% to 128.21%, driven by advancements in storage solutions and AI demand [18] - Tianfu Communication projects a net profit of 1.881 billion to 2.150 billion yuan for 2025, reflecting a growth of 40% to 60% due to the accelerating AI industry and global data center construction [19] Group 3: Regulatory and Market Responses - The European Union plans to phase out "high-risk suppliers" in critical sectors, interpreted as targeting Chinese tech firms like Huawei, which has expressed concerns over the fairness of such regulations [2] - Pinduoduo was fined 100,000 yuan for failing to report tax information as required, highlighting regulatory scrutiny on internet platform companies [4] - Zhiyu Technology announced a temporary limit on the sale of its GLM Coding Plan due to high demand and resource constraints, reducing daily sales to 20% of current levels [3]

Group 1 - The Ministry of Industry and Information Technology (MIIT) has announced the establishment of a safety monitoring platform for the operation status of new energy vehicles, effective from January 1, 2027 [4] - Xibei Catering Group has completed a new round of financing, with investors including Taizhou Xinrongtai Investment and former Ant Group CEO Hu Xiaoming, although the specific amount remains undisclosed [4][5] - The financing has increased Xibei's registered capital from 89.90 million yuan to 101.68 million yuan, marking a 13.1% increase [5] Group 2 - The price of gold jewelry in China is approaching 1500 yuan per gram, with brands like Chow Tai Fook and Lao Feng Xiang reporting significant price increases [7] - OpenAI has announced plans to expand its AI infrastructure in the U.S. to 10 gigawatts by 2029, committing to cover energy costs to prevent price hikes [12] - Nvidia's CEO Jensen Huang emphasized the rising demand for skilled tradespeople in the AI era, predicting that plumbers and electricians could earn six-figure salaries due to the infrastructure needs of AI [10] Group 3 - Apple plans to upgrade Siri into a chatbot by the second half of 2026, utilizing Google's Gemini model [10] - DeepSeek has revealed a new model, MODEL1, which is designed for efficient inference and optimized for edge devices [9] - The VCSEL chip provider Raysees Technology has completed a multi-hundred million yuan Series C financing round [20]

Real Estate Development - The Ministry of Housing and Urban-Rural Development emphasizes the importance of accelerating transformation and upgrading for high-quality real estate development, focusing on two main areas: orderly promotion of "good housing" construction and the establishment of a new model for real estate development [2] - The construction of "good housing" involves collaboration among government, enterprises, and society, with a comprehensive deployment to enhance housing quality through standards, design, materials, construction, and operation [2] - The new model for real estate development aims to ensure a smooth transition from old to new models, focusing on a mechanism that links people, housing, land, and finance [2] Real Estate Financing and Sales - The project company system will be implemented to ensure independent legal rights and responsibilities, prohibiting headquarters from misappropriating project funds before delivery [3] - A lead bank system will be introduced for real estate financing, where one bank or syndicate will be responsible for managing project funds [3] - The promotion of a "current housing sales" system aims to mitigate delivery risks, while pre-sale funds will be regulated to protect buyers' rights [3] Market Trends - The AI technology market is expected to grow significantly, with new personal AI devices emerging and the overall market scale likely to expand further between 2026 and 2027 [3] - The integration of energy and computing networks is crucial for enhancing global competitiveness, as highlighted by industry leaders [4] Mergers and Acquisitions - Energy Fuels has agreed to acquire Australian Strategic Materials for AUD 447 million (approximately USD 300.9 million), marking a significant move to secure the supply chain for rare earth elements [7] Policy Developments - The Ministry of Finance announced the establishment of duty-free shops at 41 ports, allowing residents from Macau to purchase duty-free goods [8] - New tax policies for innovative enterprises' CDRs will be implemented from January 1, 2026, to December 31, 2027, including exemptions on capital gains tax for individual investors [9] Financial Sector Updates - The People's Bank of China is focusing on modernizing the payment system and enhancing cross-border payment capabilities [10] - The National Financial Regulatory Administration has released new regulations to improve the administrative licensing process, enhancing the efficiency of market access [10] Industrial and Technological Development - The Ministry of Industry and Information Technology is promoting humanoid robot technology and aims to strengthen the ecosystem for humanoid robots [11] - A notification has been issued to automate the monitoring of computing power resources across 31 provinces by the end of 2026 [12] Economic Performance - Beijing's GDP reached CNY 5.20734 trillion in 2025, growing by 5.4% year-on-year, with the tertiary sector showing the highest growth at 5.8% [18]

Core Insights - DeepSeek has updated its FlashMLA code on GitHub, revealing the previously undisclosed "MODEL1" identifier, which may indicate a new model distinct from the existing "V32" [3][4] - The company plans to launch an "open source week" in February 2025, gradually releasing five codebases, with Flash MLA being the first project [4] - Flash MLA optimizes memory access and computation processes on Hopper GPUs, significantly enhancing the efficiency of variable-length sequence processing, particularly for large language model inference tasks [4] Company Developments - DeepSeek's upcoming AI model, DeepSeek V4, is expected to be released around the Lunar New Year in February 2025, although the timeline may vary [4] - The V4 model is an iteration of the V3 model released in December 2024, boasting advanced programming capabilities that surpass current leading models like Anthropic's Claude and OpenAI's GPT series [5] - Since January 2026, DeepSeek has published two technical papers introducing a new training method called "optimized residual connections (mHC)" and a biologically inspired "AI memory module (Engram)" [5] Industry Context - The introduction of the Engram module aims to improve knowledge retrieval and general reasoning, addressing inefficiencies in the Transformer architecture [5] - The support from Liang Wenfeng's private equity firm, which has achieved a 56.55% average return in 2025, has bolstered DeepSeek's research and development efforts [5]

Core Insights - The article discusses the emergence of a new model named "MODEL1" from DeepSeek, coinciding with the one-year anniversary of the DeepSeek-R1 release, indicating potential advancements in AI model architecture [2][6]. Group 1: Model Development - "MODEL1" has been referenced in the updated FlashMLA code on GitHub, suggesting it may represent a new model distinct from the existing "V32" architecture [2][3]. - There are differing opinions in the industry regarding whether "MODEL1" is a version 4 model or an advanced inference model, with some developers speculating it could be the ultimate version of the V3 series [2][5]. - Key technical differences between "MODEL1" and "V32" include variations in key-value (KV) cache layout, sparsity handling, and support for FP8 data format decoding, indicating targeted design for memory optimization and computational efficiency [5]. Group 2: Anticipated Release and Features - The structure of the model files suggests that "MODEL1" is nearing completion or inference deployment, awaiting final weight freezing and testing validation, which implies a forthcoming launch [5]. - There are expectations for DeepSeek to release its next flagship model, DeepSeek V4, in February, with preliminary tests indicating it may surpass other top models in programming capabilities [6]. - Recent technical papers from DeepSeek introduce new training methods and an AI memory module, hinting that these innovations may be integrated into the upcoming model [6]. Group 3: Industry Impact - The DeepSeek-R1 model has been recognized as the most praised model on Hugging Face, significantly lowering barriers in inference technology and production deployment, thus influencing the open-source strategy of major Chinese companies [9]. - Over the past year, Chinese AI models have seen increased downloads on Hugging Face, surpassing those from the U.S., indicating a shift in reliance on Chinese-developed open-source models within the global supply chain [9].

来源：深网 在R1发布一周年之际，DeepSeek 在全球AI圈再次掀起波澜。 需要指出的是，截至目前，DeepSeek 官网及微信公众号尚未披露任何关于Model1 的相关信息，其最新 一篇推送仍停留在 2025年12月1日发布的 DeepSeek-V3.2正式版公告。 在过去一年中，DeepSeek 以"小步快跑"的方式持续推进 V3 模型的迭代，重点围绕复杂推理、编程能力 和工具调用等方向进行深度优化与架构创新，同时将 R1 作为稳定基线持续赋能生态。 业界之所以猜测DeepSeek会在今天春节复刻去年R1的"核爆"，主要基于两条线索。一是有外媒称， DeepSeek预计将于2月中旬推出其下一代人工智能模型V4。 近日，DeepSeek在FlashMLA代码库更新中意外曝光了一个名为 Model1 的新模型，这一发现迅速在技 术社区引发热议。 神秘的 Model1不仅出现在代码和注释中，还拥有与 DeepSeek-V3.2 并列的独立文件。这或意味着其并 未沿用 V3 系列的参数配置或基础架构，或是一条全新的技术路径。 对此，不少网友推测这可能是DeepSeek蓄势已久、即将投向全球AI赛场的下一枚"王 ...