Core Viewpoint - The DeepSeek team has released a new academic paper focusing on optimizing inference speed for large language models (LLMs), which is crucial for the practical application of AI agents [4][5]. Group 1: Research and Innovation - The paper, co-authored with Peking University and Tsinghua University, introduces an innovative inference system called DualPath, designed to enhance the performance of LLMs under agent workloads [4]. - The DualPath system employs a "dual-path reading KV-Cache" mechanism, redistributing storage network load, resulting in an offline inference throughput increase of 1.87 times and an average increase of 1.96 times in the number of agent operations per second for online services [4][5]. Group 2: Industry Context and Expectations - The introduction of DualPath addresses the significant changes in inference workloads as LLMs evolve from simple dialogue systems to complex agent systems capable of multi-turn interactions, which can reach dozens or even hundreds of rounds [4]. - There is a growing expectation for the release of DeepSeek's next flagship model, DeepSeek V4, with various rumors about its launch timeline ranging from early February to March [6]. - Recent leaks suggest that DeepSeek is testing a V4 Lite model, codenamed "Sealion-lite," which supports a context window of 1 million tokens and native multimodal inference [6]. Group 3: Market Reactions and Concerns - Despite the technical advancements presented in the paper, there is a sentiment in the industry that such optimizations are seen as a necessity due to GPU shortages, with some viewing it as "dirty work" rather than innovative [5]. - Concerns have been raised among investment institutions that the release of the new model could lead to significant market volatility, similar to the previous year's model launch [6].

论文延续DeepSeek一贯的风格，在工程化层面将性能优化推向极致。 在业界对新一代旗舰模型DeepSeek V4的翘首期盼中，DeepSeek团队却悄然放出了一篇新的学术论文。 这篇论文由DeepSeek联合北大、清华共同撰写，将研究方向投向了决定大模型实际应用落地的关键一环——推理速度，为日益复杂的AI智能体，提供一套 高效的底层系统解决方案。 论文在引言部分提到，大模型正从单轮对话机器人和独立推理模型，快速演进为智能体系统 ——能够自主规划、调用工具，并通过多轮交互解决实际任 务。这种应用范式的转变，推动大模型推理工作负载发生重大变革：从传统的人类-大模型交互，转向人类-大模型-环境交互，交互轮次可达数十甚至数百 轮。 上下文会跨轮次累积，最终长度可能达到极值。此时模型不需要大量计算，反而需要频繁从硬盘读取历史上下文的 KV-Cache；现有系统中，只有负责预处 理的引擎会读取KV-Cache，它的网卡带宽被占满，而负责生成内容的解码引擎，网卡带宽基本闲置，导致整个系统速度被卡脖子。 因此，论文提出的DualPath，针对智能体工作负载、重新设计现代推理架构中 KV-Cache加载逻辑，解决大模型做智能 ...

今日早盘，A股继续小幅震荡，上证指数红绿间转换超10次，中证1000表现较强，连续第4日上涨，创 2017年4月以来近9年新高，深证成指、沪深300等则小幅下跌，市场成交保持平稳。 盘面上，稀有金属、人工智能、超临界发电、酒店餐饮等板块涨幅居前，玻璃玻纤、通信设备、消费电 子、航空装备等板块跌幅居前。 超临界发电项目持续推进 春节后，超临界发电概念频频走强，板块指数连续4日创历史新高。金现代早间仅约6分钟就垂直20%涨 停，豫能控股秒速涨停，连续第7日涨停，赣能股份亦秒板，连续第3日涨停，华银电力1分钟涨停，连 续第2日涨停。 近来，超临界发电领域利好不断。继去年底全球首台商用超临界二氧化碳发电机组在贵州六盘水首钢水 城钢铁（集团）有限责任公司成功商运后，贵州能源大方2×66万千瓦超超临界燃煤发电项目，也于今 年2月上旬完成全部27项前期核准及开工手续，主厂房基础浇筑全面启动。 项目2026年计划完成投资23.49亿元，计划于2027年底建成投产，年发电量约60亿千瓦时，配置新能源 指标290万千瓦，推动"火电+新能源"多能互补。 此外，中核集团50兆瓦"熔盐储能+超临界二氧化碳发电"示范项目已入选国家能源领 ...

Group 1 - AI-related stocks have seen significant gains, with Kingsoft Cloud (03896) up 9.05% to HKD 7.23, XunCe (03317) up 8.82% to HKD 89.45, Huily Technology (01860) up 6.27% to HKD 12.38, and Kingdee International (00268) up 4.76% to HKD 10.35 [1] - DeepSeek is reportedly testing the V4Lite model, codenamed "Sealion-lite," which features a context window of 1 million tokens and supports native multimodal reasoning [1] - Nomura's research indicates that the technological breakthroughs of DS-V4 will effectively break the "chip wall" and "memory wall," enabling the dual development of domestic computing hardware and AI applications, thus advancing the maturity of China's open-source large model ecosystem [1] - Industrial Securities believes that V4 is expected to be released in February, with significant potential in its application ecosystem [1] Group 2 - OpenRouter, the world's largest AI model API aggregation platform, reported that from September 9 to 15, Chinese models achieved a call volume of 41.2 trillion tokens, surpassing the 29.4 trillion tokens of U.S. models for the first time, with four domestic large models ranking among the top five globally [1]

2月，中国AI的模型调用量爆发式增长，首次超过美国。 全球最大的AI模型API聚合平台OpenRouter数据显示，9日~15日这周，中国模型以4.12万亿Token的调用量，首次超过同期美国模型的2.94万亿Token。 榜单洗牌：中国Token调用量首超美国，四款大模型霸榜 OpenRouter平台，汇聚了全球数百种大语言模型，拥有超过500万开发者用户，是目前全球最大的AI模型API聚合平台。因此，其API调用量数据被视为 洞察全球AI应用落地趋势最真实的"晴雨表"，因为它直接反映了开发者"用脚投票"的选择，体现了模型在实际应用中的受欢迎程度和竞争力。 值得注意的是，该平台的用户主要由海外开发者构成，其中美国用户占比高达47.17%，而中国开发者仅占6.01%，这使得其榜单数据更能客观反映中国AI 模型在全球范围内的真实吸引力。 16日~22日这周，中国模型的周调用量进一步冲高至5.16万亿Token，三周大涨127%，而同期美国模型调用量跌至2.7万亿Token。与此同时，全球调用量 排名前五的模型中，中国模型占据四席，这股强大的增长动能，并非依赖单一爆款产品，而是中国AI厂商集群式崛起。 Token ...

Core Insights - A new reasoning framework for agents called DualPath has been introduced, which addresses I/O bottlenecks in long-text reasoning scenarios by optimizing the speed of loading KV-Cache from external storage [1][3]. Group 1: DualPath Framework - DualPath changes the traditional Storage-to-Prefill loading mode by introducing a second path, Storage-to-Decode, allowing for more efficient data handling [3][6]. - The framework utilizes idle storage network interface card (SNIC) bandwidth from the decoding engine (DE) to read caches and employs high-speed computing networks (RDMA) to transfer data to the prefill engine (PE), achieving global pooling of storage bandwidth and dynamic load balancing [3][13]. Group 2: Performance Improvements - In tests with a production-level model of 660 billion parameters, DualPath demonstrated a remarkable increase in offline inference throughput by 1.87 times and an average increase in online service throughput by 1.96 times [3][14]. - The framework significantly optimizes first token latency (TTFT) under high load while maintaining stable token generation speed (TPOT) [5][14]. Group 3: Technical Innovations - DualPath allows KV-Cache to be loaded into the decoding engine first, which is then transmitted to the prefill engine, alleviating bandwidth pressure on the prefill side [7][9]. - The architecture includes a central scheduler that dynamically allocates tasks based on I/O pressure and computational load, preventing congestion on any single network interface or computational resource [14][18]. Group 4: Research and Development - The first author of the paper, Wu Yongtong, is a PhD student at Peking University, focusing on system software and large model infrastructure, particularly in optimizing inference systems for large-scale deployment [15][16].

【算力闭环：即将发布的DeepSeek V4"海狮轻型版"引爆国产AI产业链的"价值重估"】 2026年新春，AI领域迎来双重变奏。 一边是DeepSeek V4"海狮轻型版"即将发布，以1M超长上下文窗口惊艳业界；另一边则是路透社爆出重磅消息： DeepSeek在V4的预发布阶段，未按惯例向英伟达、AMD开放，而是将早期访问权限独家授予了华为等国内供应 商。 这一"反常"举动，将国产大 【算力闭环：即将发布的DeepSeek V4"海狮轻型版"引爆国产AI产业链的"价值重估"】 2026年新春，AI领域迎来双重变奏。 一边是DeepSeek V4"海狮轻型版"即将发布，以1M超长上下文窗口惊艳业界；另一边则是路透社爆出重磅消息： DeepSeek在V4的预发布阶段，未按惯例向英伟达、AMD开放，而是将早期访问权限独家授予了华为等国内供应 商。 这一"反常"举动，将国产大模型与国产算力的协同推到了聚光灯下。 "海狮轻型版"的命名，似乎暗示着某种轻盈与敏捷。 DeepSeek选择华为，不仅是技术适配的选择，更是战略安全的考量。 此前，DeepSeek已在昇腾平台成功完成模型迁移，通过KernelCAT工具将推理 ...

| | | 全球主要股票指数 | | | | --- | --- | --- | --- | --- | | 市场 | 名称 | 现价 | 涨跌 | 涨跌幅 | | 欧美 | 道琼斯工业平均 | 49499.2 | 17.05 | 0.03% | | | 纳斯达克 | 22878.38 | -273.7 | -1.18% | | | 标普500 | 6908.86 | -37.27 | -0.54% | | | 欧洲斯托克50 | 6161.56 | -11.76 | -0.19% | | | 英国富时100 | 10846.7 | 40.29 | 0.37% | | | 法国CAC40 | 8620.93 又 | 61.86 | 0.72% | | | College International Party of Children Company of | 25289.02 8 ogudata 3.08 | | 0.45% | | | 俄罗斯RTS 上证指数 | 1137.83 4146.63 | -13.78 -0.6 | -1.20% -0.01% | | | 深证成指 创业板指 恒生 清教 | 145 ...

【隔夜行情】 【今日新股】 【中国AI调用量首超美国 四款大模型霸榜全球前五】2月26日，全球最大的AI模型API聚合平台 OpenRouter数据显示，9日-15日这周，中国模型以4.12万亿Token的调用量，首次超过同期美国模型的 2.94万亿Token。16日-22日这周，中国模型的周调用量进一步冲高至5.16万亿Token，三周大涨127%， 而同期美国模型调用量跌至2.7万亿Token。平台调用量排名前五的模型中，有四款来自中国厂商，分别 为MiniMax的M2.5、月之暗面的Kimi K2.5、智谱的GLM-5以及DeepSeek的V3.2。这四款模型合计贡献 了Top5总调用量的85.7%。值得注意的是，该平台的用户主要由海外开发者构成，其中美国用户占比高 达47.17%，而中国开发者仅占6.01%，这使得其榜单数据更能客观反映中国AI模型在全球范围内的真实 吸引力。 (每经网) •周四（2026年2月26日），A股市场三大指数集体收涨，截至收盘，沪指报3888.60点，涨0.34%；深证 成指报12984.08点，涨0.85%；创业板指报3052.59点，涨0.70%。总体来看，个股涨多跌少 ...

hapabapa/iStock Editorial via Getty Images Listen below or on the go via Apple Podcasts and Spotify C3.ai (AI) shares fall on missed guidance, regional sales shortfalls, and workforce reductions. (00:14) Nvidia (NVDA) pops Q4 results, guidance blow past Wall Street's forecast. (01:41) DeepSeek withholds upcoming AI model from US chipmakers, including Nvidia: report. (02:34) This is an abridged transcript. C3.ai (AI) is down 23% in premarket action after posting a quarterly earnings miss and forecast ...