丨证券研究报告丨 行业研究丨点评报告丨软件与服务 [Table_Title] AI 应用正当时： Kimi 发布论文预览下一代模型 架构，模型商业化落地继续加速 报告要点 research.95579.com 3 月 16 日，月之暗面发布论文，提前预览下一代模型的关键模块——注意力残差（Attention Residuals，简称 AttnRes）。 分析师及联系人 [Table_Author] 宗建树 刘思缘 SAC：S0490520030004 SFC：BUX668 请阅读最后评级说明和重要声明 %% %% %% %% 1 软件与服务 cjzqdt11111 [Table_Title AI 应用正当时：2] Kimi 发布论文预览下一代模型 架构，模型商业化落地继续加速 [Table_Summary2] 事件描述 3 月 16 日，月之暗面发布论文，提前预览下一代模型的关键模块——注意力残差（Attention Residuals，简称 AttnRes）。 事件评论 风险提示 丨证券研究报告丨 2026-03-22 行业研究丨点评报告 [Table_Rank] 投资评级 看好丨维持 市场表现对比图(近 ...

腾讯2025年营收7517.7亿，总薪酬1307亿，人均年薪超112万。 整理 ｜徐子欣 点击上方【36氪随声听】，一键收听大公司热门新闻。听完音频记得添加进入 【我的小程序】 中哟！ 腾讯财报披露：2025年营收7517.7亿，总薪酬1307亿，人均年薪超112万 据第一财经日报，3月18日，腾讯控股（00700.HK）发布2025年第四季度及全年财报。AI持续推动公司盈利结构优化：Q4营收 1943.7亿元，同比增长13%；Non-IFRS经营利润695.2亿元，同比增长17%。全年营收7517.7亿元，同比增长14%；Non-IFRS经 营利润同比增长18%。 下周油价或重回"9元时代" 碧桂园回应大规模召回离职员工 多款知名中成感冒药被要求补全不良反应 iPhone18将首次启用三星相机 微信能折叠发图了：三张以上可合并展示 杨植麟GTC 2026演讲：首次完整披露Kimi模型技术路线图 截至2025年12月31日，腾讯集团共有员工115849名，较2024年的110558名增加超过5000人。腾讯2025年度的总薪酬成本达到人 民币1307亿元，相较于2024年的1128亿元，同比增长约15.9%。 ...

Core Viewpoint - The article emphasizes the transformative potential of AI Agents in research workflows, moving from simple question-answering to fully delegating tasks to AI, thereby enhancing research efficiency and productivity [1]. Course Features - The course focuses on teaching participants how to enable AI to perform tasks rather than just asking questions, allowing AI to engage in critical research processes [2]. - It is designed around real research scenarios, covering common and time-consuming tasks such as literature retrieval, data analysis, and paper writing, enabling immediate application in participants' research [3]. - Utilizing OpenClaw as a practical tool, the course aims to provide a comprehensive understanding of AI Agents and how to build personalized research workflows [4]. - The course promotes a workflow mindset, encouraging participants to break down tasks into manageable parts, automate processes, and create reusable personal research systems [5]. - It aims not only to improve efficiency but also to reshape competitive advantages in research by equipping participants with advanced AI skills [6]. - The course is structured to be accessible for beginners while also providing deeper insights for those with prior knowledge, ensuring all participants can effectively learn and apply the content [7]. Course Schedule and Content - The course will take place from April 3 to April 5, with multiple sessions available for those who miss the initial live class [9]. - The curriculum includes practical sessions on OpenClaw installation, integration with various models, automation of repetitive tasks, and assistance in writing and data analysis [10][11][12][13][14]. Course Outcomes - Participants will progress from merely using AI to effectively employing AI Agents in research execution [24]. - The course aims to transform frequent research tasks into reusable intelligent workflows, reducing the need to start from scratch each time [24]. - It seeks to significantly decrease repetitive labor, allowing researchers to focus on critical thinking and innovation [24]. - Participants will gain early access to next-generation research efficiency tools, establishing a competitive edge in their field [24].

Core Insights - The AGI-Next conference highlighted the current challenges and future opportunities in AI development, particularly focusing on the capabilities and limitations of large models [3][4][5]. Group 1: Key Discussions on AGI and AI Development - Zhang Bo emphasized five fundamental deficiencies in current large models, advocating for a definition of AGI that includes executable and verifiable capabilities [3]. - Yang Qiang discussed the differentiation of agents based on their ability to autonomously set and plan goals, rather than relying on human-defined parameters [3]. - Tang Jie noted that while scaling remains a valid approach, the true exploration should focus on enabling models to possess autonomous scaling capabilities [4]. Group 2: Scaling and Model Capabilities - Yang Zhilin explained that the essence of Scaling Law is to convert energy into intelligence, emphasizing the importance of efficient approaches to reach the limits of intelligence [4]. - Lin Junyang expressed optimism about the potential for Chinese teams to achieve global leadership in AI within the next 3-5 years, estimating a 20% probability of success [4]. - Yao Shunyu highlighted the differentiation between vertical integration and layered model applications, suggesting that model companies may not necessarily excel in application development [4]. Group 3: Future Directions and Challenges - The discussion pointed out that the path from scaling to genuine generalization capabilities remains a core challenge for AI models [12][14]. - The need for models to develop memory and continuous learning structures akin to human cognition was identified as a critical area for future research [35][36]. - The exploration of self-reflection and self-awareness capabilities in AI models was deemed a significant yet controversial topic within the academic community [36][47]. Group 4: Technical Innovations and Model Architecture - The introduction of new optimization techniques, such as the Muon optimizer, was highlighted as a means to enhance token efficiency and overall model performance [55][58]. - The development of the Kimi Linear architecture aims to improve linear attention mechanisms, making them more effective for long-context tasks [64]. - The integration of diverse data sources and the enhancement of model architectures are seen as essential for achieving better agent capabilities in AI [67].

Core Viewpoint - The article discusses the unexpected shift in the U.S. tech landscape, where many American startups are increasingly adopting Chinese open-source AI models despite previous restrictions and concerns about China's AI development [2][10][24]. Group 1: U.S. Companies' Adoption of Chinese AI Models - Over half of U.S. startups are now choosing Chinese open-source AI models as their primary development tools, indicating a significant change in preference [4][10]. - Companies like Perplexity and Airbnb are openly utilizing Chinese models, with Airbnb's CEO stating their AI customer service system heavily relies on Alibaba's Qwen model [6][10]. - The cost-effectiveness of Chinese models is a major factor, with one U.S. entrepreneur noting a switch from a closed-source model that cost $400,000 annually to Qwen, which significantly reduced expenses [10][12]. Group 2: Advantages of Open-Source Models - The annual cost of closed-source models exceeds $1,000 per user, while Chinese open-source models are nearly free, providing a substantial financial incentive for companies [12]. - Open-source models offer greater control and transparency, allowing companies to modify the code as needed without the risk of sudden changes in service terms, as experienced with ChatGPT [12][14]. - The shift from closed to open-source models reflects market dynamics, where companies prioritize economic and security considerations [14][16]. Group 3: Impact of U.S. Restrictions on Chinese AI Development - U.S. restrictions on high-end GPU supplies forced Chinese teams to innovate and optimize algorithms to achieve better performance with limited resources, exemplified by the DeepSeek team [18][20]. - Chinese models are evolving from mere tools to essential infrastructure, similar to the Android system, with millions of developers building applications on these platforms [22][28]. - The competitive edge of Chinese open-source models lies in their low cost, high efficiency, and freedom, challenging the notion that technological progress can be stifled by restrictions [26][29].

Core Insights - Huang Renxun, the founder and CEO of NVIDIA, emphasized the importance of the Chinese market, noting it as the second-largest technology market globally with rapid growth [5][18] - NVIDIA's H20 GPU has been reintroduced to the Chinese market, although rebuilding the supply chain will take time [6][38] - Huang highlighted the advanced level of computer science and AI talent in China, stating that about 50% of AI researchers globally are based in China [5][31] Group 1 - Huang Renxun's visit to China this year was motivated by invitations, reflecting the significance of the Chinese market for NVIDIA [5][6] - The RTX PRO product is designed for digital twin applications, which aligns well with China's advancements in robotics and smart factories [20][30] - Huang expressed optimism about the collaboration with Chinese companies, citing a long history of partnerships with major firms like Tencent and Alibaba [11][18] Group 2 - Huang acknowledged the challenges posed by global trade policies and tariffs, stating that NVIDIA must adapt to these changes [6][16] - The company is committed to investing in the Chinese market to keep pace with competitors who are also increasing their investments [17][18] - Huang praised the innovation and craftsmanship present in China, indicating that NVIDIA's products are powering many innovative Chinese enterprises [6][19] Group 3 - Huang noted that the education system in China has produced a significant number of top AI researchers, contributing to the country's strong position in AI development [5][31] - He mentioned the importance of open-source AI models like DeepSeek, which have gained global traction and are being utilized across various industries [63][42] - Huang emphasized the need for companies to focus on creating valuable products and technologies that can impact the world positively [16][70]

Group 1 - The article discusses how AI tools can automate the generation of annual business analysis reports, enhancing efficiency and maintaining analytical depth comparable to manual writing [1][9] - Data preparation involves integrating multi-source data from ERP, CRM, and financial systems, utilizing AI tools for data cleaning and standardization [3][4] - Key performance indicators (KPIs) are selected for analysis, such as revenue growth rate, gross margin, and net cash flow, with AI tools generating comparative metrics [4] Group 2 - Various AI tools are recommended, including general-purpose tools like GPT-3/4 for text generation and DeepSeek for data modeling, as well as specialized tools like Quill for financial reporting [4][5] - The report generation process is template-driven, allowing users to upload cleaned data and select preset templates for automatic report creation [4][5] - Manual proofreading and optimization are essential, focusing on data accuracy checks and logical coherence adjustments to ensure the quality of AI-generated reports [7][8] Group 3 - Typical application scenarios include financial analysis modules that automatically generate balance sheets and profit and loss statements, as well as market trend forecasting [6][8] - Data security is emphasized, recommending local deployment of AI tools to protect sensitive business data, along with originality checks for AI-generated content [6] - The article concludes that companies can improve report writing efficiency by over 60% while ensuring depth of analysis, with future advancements expected in fully automated report generation [9]