撰 文 / 周 洲 设 计 / 琚 佳 75%。 这是芯片龙头英伟达2025财年的毛利率，堪称暴利。 2月26日，英伟达公布了2025财年及第四季度财报，并就财报召开了电话会议。 得益于近两年爆炸式增长的AI领域，数据中心业务为英伟达贡献了约90%的收入。 但中国初创公司DeepSeek在1月20日发布的低成本R1模型，打破了美国主导的巨额投入型AI模式， 也使得英伟达的股票遭遇了美国历史上最大的单日跌幅，市值蒸发近6000亿美元。 英伟达创始人兼首席执行官黄仁勋在随后的财务电话会议上，对DeepSeek点了赞。 他认为，DeepSeek不是威胁，反而能助力英伟达的芯片大卖。 他的理由是，过去几年对人工智能计算的需求仍然"才刚刚开始"，尽管最近开发了新的低成本人工 智能，但需求仍将继续增长。 全年营收翻倍，盈利增幅放缓 当地时间2月26日美股盘后，英伟达公布了2025财年第四财季和2025财年的财务业绩。 财报显示，英伟达在2025财年第四财季实现营收393.31亿美元，较去年同期增长78%，净利润为 220.66亿美元，同比上升72%。 | | | GAAP | | | | | --- | --- | -- ...

"当 AI 足够强大后，开源还是不是一个好选择？" 整理丨刘倩 程曼祺 嘉宾丨美国西北大学 MLL Lab 博士王子涵 ▲扫描上图中的二维码，可收听播客。《晚点聊 LateTalk》#102 期节目。欢迎在小宇宙、喜马拉雅、苹果 Podcast 等渠道关注、收听我们。 《晚点聊 LateTalk》是《晚点 LatePost》 推出的播客节目。"最一手的商业、科技访谈，最真实的从业者思考。" 这是《晚点 LatePost》 「开源对话」系列的第 2 篇。该系列将收录与开源相关的访谈与讨论。系列文章见文末的合集#开源对话。 上周五，DeepSeek 在官方 Twitter 上预告了下一周会连续 5 天开源 5 个代码库，进入 "open-source week"开源周。 目前 DeepSeek 已放出的 4 个库，主要涉及 DeepSeek-V3/R1 相关的训练与推理代码 。 这是比发布技术报告和开源模型权重更深度的开源。 有了训练和推理 工具，开发者才能更好地在自己的系统里，实现 DeepSeek 系列模型的高效表现。 (注：所有 4 个库和后续开源可见 DeepSeek GitHub 中的 Open-Inf ...

"真正的大模型开源，应该把数据集也开源。" 文丨贺乾明 编辑丨宋玮 过去两个月，DeepSeek 重塑全球大模型格局，也扭转了整个行业对开源的理解。 OpenAI 反思走向闭源是 "站在历史错误的一边"，百度、MiniMax、阶跃星辰等原本闭源的公司转向开源。 "如果在以前，一个拿几亿美金融资的公司说自己要开源，估计投资人会吐血。" 一位科技投资人说。 DeepSeek 还在加大开源力度。这周，DeepSeek 计划开源 5 个训练、推理大模型相关的代码库——而大多数开源模型的公司还停留 在开放模型权重层面。 到底该怎么看待 DeepSeek 的开源？它对大模型开源社区意味着什么？为什么不同公司选择不同的开源策略？选择开源对一家商业 公司到底意味着什么？ 近期，我们访谈了中国开源先驱章文嵩。他 1995 年读硕士期间接触到开源，那时中国刚通互联网不久，不少 DeepSeek 的研究者还 没有出生。 1998 年，章文嵩在国防科大读博期间开源了 LVS（Linux 虚拟服务器）软件，这个均衡服务器访问流量、避免宕机的系统，是中国 最早在全球科技行业扩散的开源项目，如今是互联网基础设施的组件。 "几乎所有的互联网 ...

Core Viewpoint - The article discusses the collaboration between 36Kr and DeepSeek, which aims to revolutionize the production of financing reports using AI technology, allowing users to receive professional and readable reports in just half an hour [2][3][7]. Group 1: AI Integration and Efficiency - 36Kr has integrated DeepSeek to create a new model for financing report production characterized by high efficiency and cost-effectiveness [2][3]. - DeepSeek has achieved over 110 million downloads and a peak of nearly 97 million weekly active users, showcasing its unique cognitive capabilities compared to other AI writing tools [3]. - The collaboration will enable startups to generate appealing narratives for investors at a low cost, helping them gain visibility in a competitive market [3][4]. Group 2: User Experience and Process - Users can access the "Seek Report" page on the 36Kr app, where an AI assistant guides them to fill in necessary information for report generation [4]. - After generation, reports undergo a manual review process before being published in the "Self-Service Reporting" section, leveraging 36Kr's platform for exposure [4][5]. - The AI system not only generates content but also matches questions to the company's reporting type, ensuring accurate and credible information through data cross-referencing [6]. Group 3: Market Impact and Vision - 36Kr aims to serve as a bridge between entrepreneurs and investors, enhancing the visibility of innovative projects through efficient content output and industry activities [7]. - The partnership with DeepSeek positions 36Kr as a participant and promoter in the wave of technological innovation in China, supporting the growth of tech companies [7].

Core Viewpoint - Tencent has launched its new foundational model, Hunyuan TurboS, which aims to enhance its competitive edge in the rapidly evolving large model sector [1][4]. Model Architecture and Cost Reduction - Hunyuan TurboS utilizes an innovative Hybrid-Mamba-Transformer architecture, which effectively reduces the computational complexity and cache usage compared to traditional Transformer structures, leading to lower training and inference costs [4][5]. - The Mamba architecture, based on State Space Model (SSM), introduces a selective mechanism that allows efficient processing of long sequence data, addressing the high costs associated with training and inference of long texts [4][5][6]. Performance Metrics - Hunyuan TurboS has been benchmarked against other models, achieving notable scores in various categories such as MMLU (89.5), GPQA-diamond (57.5), and HumanEval (91.0), indicating its strong performance in knowledge and reasoning tasks [7]. Pricing Strategy - The pricing for Hunyuan TurboS has significantly decreased, with input costs set at 0.8 yuan per million tokens and output costs at 2 yuan per million tokens, making it more accessible compared to its predecessor [8]. Market Response and Product Integration - Following the launch of Hunyuan TurboS, Tencent's AI assistant, Tencent Yuanbao, has rapidly gained popularity, surpassing Doubao in downloads and reaching the second position in the Apple Store's free app rankings in China [14][15]. - The integration of Hunyuan TurboS into Tencent Yuanbao has led to multiple significant updates, enhancing its capabilities and user experience [16]. Stock Market Reaction - Tencent's aggressive shift towards AI has positively impacted its stock price, which reached a high of 522 HKD, the highest since August 2021, before settling at 495 HKD [21].

Group 1 - DeepSeek announced a significant reduction in API calling prices during off-peak hours, with R1 and V3 model APIs seeing price cuts of 75% and 50% respectively [1] - The off-peak hours defined by DeepSeek cover the daytime in Europe and the US, indicating a strategic pricing approach to attract developers [1] - This pricing strategy follows a trend initiated last year, which sparked a price war in the AI model market, particularly after the release of DeepSeek's V2 model [1] Group 2 - The recent price cuts by DeepSeek have caused significant reactions in both domestic and international AI industries, highlighting the competitive landscape [1] - Following the launch of its AI assistant, DeepSeek's pricing strategy has prompted responses from competitors like OpenAI and Google, who have also adjusted their pricing [1]

Core Viewpoint - The open-sourcing of DeepSeek is creating significant opportunities for mid-sized AI companies and domestic chip manufacturers, while posing challenges for established large model companies known as the "six little tigers" [1][4][8]. Group 1: Impact of DeepSeek Open-Sourcing - Many mid-sized private enterprises are rapidly transitioning to DeepSeek's base model, with over half of existing clients making the switch [1]. - The open-sourcing initiative has sparked a wave of enthusiasm in AI application entrepreneurship, leading to a twofold increase in collaboration requests for domestic chip companies [1]. - The "open-source week" plan by DeepSeek, which began on February 21, aims to share several code repositories, enhancing transparency and innovation in AI [3]. Group 2: Reactions from Industry Players - Internal debates are ongoing among the "six little tigers" regarding the implications of open-sourcing, with concerns that it could disrupt their business models [2]. - The open-source trend has prompted even traditionally closed-source companies like Baidu to consider open-sourcing their models [3]. - Industry experts suggest that while DeepSeek's innovations benefit application and chip companies, base model vendors face significant challenges [3][7]. Group 3: Market Dynamics and Future Prospects - The open-sourcing of DeepSeek is expected to benefit hardware and chip manufacturers, allowing them to engage more in training and inference businesses [7]. - The algorithms and code optimizations shared during the open-source week are designed to maximize GPU performance, enabling smaller developers to build high-performance models at lower costs [7]. - Despite the advantages, many companies may struggle to implement DeepSeek's offerings without additional support from service layer companies [7][8]. Group 4: Broader Implications - The open-source movement initiated by DeepSeek is seen as a catalyst for a broader shift in the AI ecosystem, potentially leading to a more collaborative environment [10]. - The participation of DeepSeek in major developer conferences indicates a strategic move to solidify its position in the market and expand its influence [10]. - As more companies integrate DeepSeek, questions arise regarding the commercialization and sustainability of its services [10].

Core Viewpoint - The article introduces a new ranking method called Prompt-to-Leaderboard (P2L) that allows users to input any prompt and receive real-time rankings of large models, identifying the most suitable model for that prompt [1][10]. Group 1: P2L Ranking Mechanism - P2L ranks models based on their performance in response to specific prompts, enabling users to find the model that best addresses their needs [1][10]. - The ranking is dynamic, with models being evaluated in real-time as prompts are entered, showcasing their scores and relative performance [5][9]. - The system highlights the differences in model performance based on the nature of the prompt, such as the impact of content restrictions on rankings [7][10]. Group 2: Model Performance Examples - For a mathematical prompt, the model "03-mini-high" achieved the highest score of 1228, demonstrating its effectiveness in handling numerical tasks [5]. - In a prompt requiring HTML, CSS, and JS code for a 3D Earth, the model "Nous-Hermes-2-Mixtral-8x7B-DPO" scored 1257, indicating its proficiency in programming tasks [9]. - The rankings for prompts related to sensitive or inappropriate content showed that less restricted models performed better, while those with strict guidelines ranked lower [7][10]. Group 3: Additional Features and User Interaction - The platform offers a "P2L Router" feature that automatically selects the best model to respond to user prompts, enhancing user convenience [22][24]. - Users can explore various categories and subcategories to compare model performance across different tasks, providing a comprehensive view of model capabilities [18][20]. - The system also allows for user feedback and interaction, raising questions about the reliability and optimization of the ranking mechanism [25][26]. Group 4: Methodology and Evaluation - P2L utilizes a Bradley-Terry (BT) model to predict user preferences based on specific prompts, aiming to provide a more accurate ranking than traditional global rankings [29][30]. - The methodology focuses on the impact of prompts on model performance, allowing for tailored evaluations that reflect real-world usage scenarios [31][32]. - Experimental results indicate that P2L outperforms traditional ranking methods, particularly as the scale of models and datasets increases [35].

Nvidia has signalled no drop in demand for its flagship chips among big artificial intelligence (AI) spenders despite the low-cost challenge posed by Chinese rival DeepSeek.The leading AI chipmaker said it expected Blackwell sales to continue to grow after its latest earnings beat market expectations. Nvidia forecast revenue of around $43bn (£34bn) for its first quarter after achieving a figure of $39.3bn (£31bn) over its last three months - up 12% from the previous quarter and 78% from one year ago.Just a ...

This article has been reviewed according to Science X's editorial process and policies . Editors have highlighted the following attributes while ensuring the content's credibility: Nvidia chief Jensen Huang says the Silicon Valley chip titan has successfully ramped up production of its new Blackwell processor that power artificial intelligence in data centers. Nvidia on Wednesday said it finished its fiscal year with record high revenue of $130.5 billion, driven by demand for its chips to power artificial ...