中信建投（601066）认为，效率优化下，可盈利AI商业模型已然跑通。DeepSeek连续开源在训练 和推理效率方面的多项技术，公布的成本利润率细节更是构建了可盈利的商业模型。该机构对 DeepSeek实际利润率进行测算，如果DeepSeek不准备冗余算力，仍能在API调用付费率20%，V3调用占 比50%的假设下实现21%的成本利润率水平；如果API调用付费率进一步提升至50%，成本利润率水平 将进一步提升至51%。 南向资金年内净流入超4000亿港元，如何一键配置港股科技公司？公开信息显示，恒生科技指数 ETF（场内：513180；联接A/C：013402/013403）标的指数囊括30家港股科技龙头，软硬科技兼备，成 分股深度聚焦AI产业链的上中下游，其中阿里、腾讯、小米、美团、中芯国际等有望成为中国科技 股"七巨头"。恒生科技指数代表了中国AI核心资产，长期有望高景气。 3月26日早盘，港股三大指数小幅高开，恒生指数涨0.5%，报23460.31点，恒生科指涨0.36%，国企 指数涨0.31%。盘面上，科网股普涨，汽车股普涨，有色金属集体高开。开盘后，恒生科技指数ETF （513180）跟随指数小幅上 ...

再次打破传统！DeepSeek发布更新，可以直接在消 费级硬件上运行 同时，新版V3模型基于R1的写作水平进行了进一步优化，同时特别提升了中长篇文本创作的内容质 量。新版V3模型可以在联网搜索场景下，对于报告生成类指令输出内容更为详实准确、排版更加清晰 美观的结果。此外，在工具调用、角色扮演、问答闲聊等方面，新模型也得到了提升。 值得一提的是，业内的早期测试证实，该模型可以直接在消费级硬件上运行。 据报道，AI研究员Awni Hannun表示，新的DeepSeek-V3模型可以在配备M3 Ultra芯片的苹果电脑上，以 每秒20个token的速度运行。这打破了业界关于人工智能模型能力与本地化运行或冲突的早前共识，也 意味着数据中心并不是大模型的必要搭配。 本文系观察者网独家稿件，未经授权，不得转载。 据官网公告，DeepSeek V3模型已完成小版本升级，目前版本号DeepSeek-V3-0324，用户登录官方网 页、APP、小程序进入对话界面后，关闭"深度思考"即可体验。API接口和使用方式保持不变。"如非复 杂推理任务，建议使用新版本V3模型，即刻享受速度更加流畅、效果全面提升的对话体验。" 新版V3模型 ...

DeepSeek V3再次震撼硅谷，中美AI差距突然缩至3 个月！ 中国人工智能初创企业DeepSeek近日发布其最新大语言模型DeepSeek-V3-0324，以全面升级的技术架构 向OpenAI、Anthropic等美国AI领军企业发起挑战。这一跨越式进展不仅彰显中国在人工智能领域的雄 心，更将中美AI竞赛推向新高度。 01.AI创始人、前谷歌中国总裁李开复表示，DeepSeek通过算法创新和高效利用国产硬件，显著缩小了 与美国领导者如OpenAI的技术差距。这一进展表明，中国在核心AI技术上仅落后美国三个月，甚至在 某些领域已处于领先地位。李开复在接受路透社采访时表示： 相较于前代产品，V3版本在以下维度实现显著提升： 硅谷企业已提高警惕，Anthropic在其最新融资文件中将中国AI技术列为"最大战略威胁"。与此同时，资 本市场开始调整布局，红杉资本等投资机构已设立专项基金，加大对本土AI项目的投入。在商业化应 "此前我认为差距在六到九个月，且全面落后。而现在，我认为在部分核心技术领域仅落后 三个月，但在某些特定领域已实现领先。" 今年早些时候，DeepSeek发布了一款基于性能较低芯片训练的AI推理 ...

路透社今年2月底引述3名知情人士的说法宣称，DeepSeek原计划在今年5月初发布R2，但现在希望尽早 推出，具体时间尚未透露。此外，DeepSeek希望新模型在代码生成和多语言推理方面的表现进一步提 升。不过，外媒的相关传言并没有得到DeepSeek公司的证实与回应。 沈阳表示，DeepSeek-V3-0324的推出进一步凸显中国AI企业在技术与成本上的竞争力。美国对华GPU出 口限制可能促使中国企业加速国产硬件适配，同时其开源模式或引发西方厂商的连锁动作，例如推出更 强闭源模型。2025年可能是中美AI竞争的分水岭。 沈阳认为，在OpenAI公司的GPT大模型要把通用大模型和推理大模型融合在一起的背景下，外界关注 包括DeepSeek在内的中国头部大模型是不是最终也会出现这种合并的趋势。"这种可能是存在的，因为 对于用户来说，并不关心大模型在回应自身问题时用的是什么类型的模型，更关心大模型能不能给出更 为智能、合理的参考答案。" DeepSeek移动端页面 图源：视觉中国 在回答《环球时报》记者有关DeepSeek-V3新版本有哪些能力提升时，DeepSeek表示，一是新版本代码 能力显著提升，接近Cla ...

Core Insights - DeepSeek has officially announced the release of its V3 model, which has garnered significant attention for its enhancements in inference, front-end development, Chinese writing, and search capabilities. This model is now recognized as the highest-scoring non-inference model, surpassing competitors like xAI's Grok3 and OpenAI's GPT-4.5 [1][4] Group 1: Model Enhancements - The V3 model represents a substantial upgrade over the previous R1 model, utilizing reinforcement learning techniques to significantly improve performance in inference tasks [6] - In code-related tasks, the V3 model generates more usable and visually appealing code, exemplified by a program simulating multiple balls in motion with adjustable parameters [6] - The model has improved the quality of mid to long-form Chinese text creation and provides more detailed and well-formatted outputs for report generation in online search scenarios [6] Group 2: Performance Metrics - The V3 model has achieved a 7% increase in intelligence index, leading all other non-inference models, although it still trails behind DeepSeek's own inference model R1 and other inference models from OpenAI, Anthropic, and Alibaba [7] - Despite being a non-inference model, the V3 model's ability to provide immediate responses makes it particularly useful in scenarios sensitive to latency [7] Group 3: Developer Feedback - Developers have reported significant improvements with the V3 model, noting its ability to surpass the R1 model and even Claude-3.7 in practical coding tests, demonstrating visible advancements in physical motion simulation [7] - An overseas developer successfully created a website and wrote over 800 lines of code without any errors using the new model, highlighting the competitive pressure open-source models are placing on larger tech companies [8]

新版DeepSeek-V3登顶非推理模型榜单！每经记者实测编程能力，R2模型也要 来了？ 每经记者 岳楚鹏 每经编辑 兰素英 北京时间3月24日晚间，DeepSeek悄然将DeepSeek-V3模型的最新版本上传到了开源平台HuggingFace。 新模型的版本号为DeepSeek-V3-0324，参数为6850亿，较初代V3版本的6710亿有小幅增长。 尽管DeepSeek十分低调，但还是有不少人在第一时间就注意到了这一更新，并对其进行了测试。 根据社区测试反馈，DeepSeek-V3-0324最明显的变化是编程能力得到了极大的提升。众多开发者基于对新模型的综合体验判断，新模型的编程能力已经接近 目前最强编程模型Claude 3.7 Sonnet。 3月25日，专业AI模型评测机构Artificial Analysis发布的最新排名显示，新版V3在基准测试中较老版V3跃升了7位，排名所有非推理模型中的第一名。 《每日经济新闻》记者实测后发现，DeepSeek-V3-0324的编程能力确实强大，但仍会出现幻觉问题。 有外媒推测："V3新版本的推出时机和特点强烈表明，它将成为DeepSeek-R2的基础，后者是 ...

Core Insights - A high school student, Adi Singh, has developed a new AI evaluation benchmark called MC-Bench, utilizing the game Minecraft to assess AI models' capabilities in a more intuitive manner [1][2][10] - Traditional standardized tests often give AI models an unfair advantage, as they are optimized for specific tasks, leading to discrepancies in real-world performance [2][8] - MC-Bench allows users to vote on AI-generated architectural designs in Minecraft, providing a crowdsourced method for evaluating AI performance [5][9] Group 1: MC-Bench Overview - MC-Bench is designed to evaluate AI models by having them create structures in Minecraft based on user prompts, such as "a crystal-clear wine glass filled with deep red wine" [2][5] - The evaluation process involves user voting to select the best creations, with results revealed only after voting concludes [5][10] - The project has garnered attention from major AI companies like OpenAI, Google, and Anthropic, which provide computational resources but are not officially collaborating [10][13] Group 2: Advantages of Game-Based Evaluation - Minecraft serves as a familiar and visually engaging platform, making it easier for the general public to understand and participate in AI assessments [7][8] - The game environment allows for a controlled testing space, enabling the evaluation of AI's reasoning and planning abilities in a safe manner [7][8] - Game-based assessments can simulate real-world complexities, test AI's decision-making skills, and provide a repeatable environment for comparison [7][8] Group 3: Current Status and Future Plans - As of now, MC-Bench primarily tests basic construction abilities of AI models, tracking their progress since the GPT-3 era [10][16] - Future plans include expanding the benchmark to more complex tasks that require long-term planning and goal-oriented actions [10][16] - The leaderboard of MC-Bench shows that Claude 3.7 Sonnet ranks first, while DeepSeek-R1 is currently in third place, indicating the platform's effectiveness in reflecting user experiences with these models [14][16]

Core Insights - The Chinese AI startup landscape is undergoing significant changes as companies adjust their business models in response to the success of DeepSeek, which has led to a concentration of market power among a few leading firms [1][2][3] Group 1: Business Model Adjustments - Many Chinese AI startups are shifting resources towards application development rather than foundational model development due to the competitive pressure from DeepSeek [1] - Zero One Everything, founded by former Google China head Kai-Fu Lee, is transitioning its business to align with what it calls the "DeepSeek era," ceasing pre-training of large language models by the end of 2024 [1] - The company announced it will offer enterprise-level DeepSeek deployment customization solutions, leveraging its expertise in hybrid expert models [1] Group 2: Funding and Investment - The startup Moonlight is reducing its marketing budget for its chatbot Kimi and focusing on model training to enhance performance, having raised over $1.3 billion (approximately 9.4 billion RMB) in funding in 2024 [2] - Alibaba has shown interest in acquiring Moonlight, having invested $800 million, which includes rights for future purchase, although recent shifts in focus may lower the likelihood of this acquisition [2] Group 3: Sector Focus Changes - Baichuan Intelligence is pivoting towards the healthcare sector, having dissolved its financial AI sales team to concentrate on developing AI technologies for medical diagnostics [3] - Zhipu AI, founded by renowned computer scientist Tang Jie, is exploring multiple business avenues and aims for an IPO by the end of 2025, although DeepSeek's growth may impact this plan [3] - Zhipu AI reported sales of 300 million RMB in 2024, with losses amounting to 2 billion RMB [3]

后DeepSeek时代：六小虎向左，BAT向右 DeepSeek犹如一颗投入平静湖面的巨石，在AI行业掀起了滔天的波澜，甚至可以夸张点说，其直接改写了国内大模型的竞争规则。 DeepSeek给AI大模型行业，免费赠送了一波国民级别的市场教育，却也平等地在先行者们头上，悬起了一把达摩克利斯之剑。 其中，AI"六小虎"之中的智谱就是一个缩影，智谱脱胎于清华大学知识工程研究室，素来有"国家队"之称。然而就在最近开始频频出现融资动作，10天之 内补充弹药达15亿人民币；可与此同时，组织震荡颇有加剧之势，从一线团队到高管大牛皆有波及。冰火两重天的态势，可谓是目前除了DeepSeek之 外，大多数大模型从业者们，真实写照的一个缩影。 2024年底，智谱曾以200亿元的估值，完成一轮30亿元人民币的融资，在这之后，包括杭州城投、上乘资本、华发集团等国资背景的资方快马加鞭地赶到 为其注资。 不过，也有风投人士对「新熵」分析，DeepSeek的横空出世还是对智谱的估值造成了一定负面影响，快速拿钱也可能是为了抢下已经出现上涨瓶颈的相 对高价。 与大开现金粮仓之门形成反差的是，智谱在团队规模和对外投资上呈现出收缩之势。高峰期阶段的 ...

Core Insights - The CEO of Zero One Technology, Kai-Fu Lee, stated that the gap between China and the U.S. in AI development has narrowed to just three months in certain areas due to advancements by companies like DeepSeek [3] - Lee emphasized that the rise of DeepSeek indicates China's leading position in infrastructure and software engineering [3] - He noted that U.S. semiconductor sanctions act as a "double-edged sword," presenting challenges but also driving innovation within Chinese companies [3] Company Developments - Zero One Technology is focusing on practical AI applications, specifically software solutions that help clients better deploy foundational models [4] - The company recently launched an all-in-one AI work platform called "Wanzhi," aimed at assisting enterprises in deploying AI technology [4] - Zero One Technology has begun generating revenue and anticipates significant growth in income, projecting to reach several times last year's revenue of $15 million by 2025 [4]