Table_First|Table_Summary 华福证券 固收定期研究 2025 年 7 月 6 日 银行/红利板块的持续性如何看 ——转债周策略 ➢ 银行/红利板块的持续性如何看 近期银行板块表现较好，从中期的角度看银行在内的红利板块表现也较稳定。我们判断 当前红利板块是较有性价比的赛道，背后的主要原因在于红利板块保险资金持续流入的 确定性较强。 22年至今，保险资金运用余额的规模持续稳健增长，主因或在于保费的规模扩容；另一 方面，保险资金运用余额中，股票投资规模也呈现震荡上行的趋势，将股票投资增速按 总规模和股票投资仓位进一步拆分后，我们看到总规模的增速相对稳定，而仓位增速波 动相对较大，23年Q2-Q4仓位增速为负，24年Q2至今仓位回归正增速。 我们将沪深300的3月区间涨跌幅平均值与仓位增速进行相关性分析，可以看到保险股票 投资仓位和股票市场赚钱效应相关性较强。当前股票市场情绪有所回暖，股指震荡上涨 的背景下，我们判断中期保险股票投资仓位或再度回升。因此对红利板块和其中的银行 板块形成资金流入增量。 从转债投资的角度，我们可基于市值、盈利稳定性、股息率等角度筛选符合险资投资偏 好的红利转债。另外 ...

Group 1 - The report addresses the differences between dexterous hands and grippers, emphasizing the importance of dexterous hands in robotic applications [1] - Dexterous hands have four key modules: drive, deceleration, transmission, and perception, which are crucial for the commercialization of robots [1] - In industrial applications, grippers are favored for their stability and cost-effectiveness, while dexterous hands offer greater adaptability and flexibility for complex tasks [1] Group 2 - The complexity of dexterous hands is expected to drive demand for various components, with a notable increase in degrees of freedom showcased at the "werobot" event [2] - The application of micro-screws and tendon ropes is expanding among manufacturers like Tesla for more efficient transmission solutions [2] - The market for dexterous hands is projected to reach 900 million yuan by 2025 and 37.6 billion yuan by 2030, with a compound annual growth rate of 110% [2] Group 3 - The competitive landscape for dexterous hands includes three main types of players: self-researching robot manufacturers, startups focused on dexterous hands, and companies extending from other industries into dexterous hand components [3] - Companies with quality products related to dexterous hands include Longsheng Technology, Xiangxin Technology, Wuzhou Xinchun, Nanshan Zhishang, Hengshuai Co., Guangyang Co., and Helit Technology [3]

题图｜AI生成 智能体的浪潮正在改写AI产业的主导逻辑。 过去一年，从内容生成走向任务执行， 智能体成为大模型能力的新出口，也成为独立模型厂商和互 联网大厂争相布局的新赛点 。 在这场变局中， 大模型厂商面临艰难抉择：是继续作为通用能力的提供者，退居底层？还是向上 走，构建平台、直达应用？ 商业化困局与技术路径选择交织下，哪些玩家能留在牌桌？ 近期，《虎嗅·AI无悖论》节目特别邀请智谱高级商业副总裁、前字节跳动飞书&Lark全球首席商业 官吴玮杰，以及企业知识开源计划创始人、前波士顿咨询Platinion董事总经理、前IBM咨询全球合伙 人陈果进行了探讨，请他们分享了智能体技术冲击下，模型厂商的战略近况及行业相关的深度思考： 智能体的普及，是在放大大模型的基础设施作用，还是正在削弱模型厂商的核心价值？ 本期主持人为资深媒体人、热AInext主理人陈庆春。 以下是交流实录，有删编： 智能体是什么，谁在制造？ 主持人：智能体（Agent）到底是什么？它和大模型是什么关系？ 吴玮杰： 我会把大模型当做是一个图书馆。现在很多chat类的应用，更像图书管理员，当它接收到 一个用户的指令，就会去图书馆里查阅各种图书，给出 ...

如果您有任何科研辅导需求，欢迎联系我们！ 自驾方向： 大模型、VLA、端到端自动驾驶、3DGS、BEV感知、目标跟踪、毫米波雷达视觉融合、激光视觉融合、 多传感器标定、多传感器融合、车道线检测、在线地图、轨迹预测、世界模型、3D目标检测、Occupancy、高性能计 算、NeRF、语义分割、决策规划等。 具身方向： VLA、视觉语言导航、端到端、强化学习、Diffusion Policy、sim2real、具身交互、抓取点预测与位姿估 计、机器人决策规划、运动规划、3DGS、SLAM、触觉感知、双足/四足机器人、遥控操作、零样本学习等； 3D视觉 相关： 点云处理、3DGS、SLAM等； 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近15个 方向 学习 路线 今年的CVPR和ICCV工作陆续放出，从今年的2大顶会来看热点研究方向，主要集中在四个部分：通用cv、自动驾驶 相关、具身相关、3D视觉相关。下面是4个领域中细分的子方向。 计算机视觉与图像：diffusion、图像质量评估、半监督学习、零样本、开放世界检测等； 自动驾驶相关，集中在端到端、闭环仿真3DGS、多模态大模型、扩 ...

下周又要看关税后续了 90天到了。 目前越南的20% ，周五市场解读为利好。 不知道后续nv卡 等有啥招数。 唯一不变的是 三四季度 一批asic卡要面世了。 国内卡荒将得到一定缓解。 但是nv的政策，现在也不知道走向何处？？？ 1）Anthropic年化收入达到40亿美元，较年初增长近四倍； 2）Cursor通过引入Anthropic的高管增强市场竞争力； xAI控制台新增Grok4及Grok4Code引用， 全球第二Anthropic年化收入已达40亿美元，较年初增长近4倍、与 Cursor 竞争加剧 海外基建 落地： 阿里云将设立首个AI全球能力中心，并在马来西亚、菲律宾新增数据中心 阿里云将设立首个AI全球能力中心，计划联合超过1000家企业打造10余个行业AI示范项目，并与全球 120多所高校合作，每年培养10万名AI人才。 接下来还是 asic逻辑和稳定币还是不受关税影响。其它不知道。 1）Grok4作为xAI的旗舰模型，专注于自然语言处理、数学推理和综合推理能力的提升； 2）Grok4Code专为编程优化，计划与代码编辑器无缝整合，提高开发效率； 3）xAI通过API提供Grok4访问权限，未来 ...

本周专题：特斯拉 Q2 交付下滑 小鹏 G7 正式上市 汽车 2025 年 07 月 06 日 特斯拉 Q2 交付下滑 小鹏 G7 正式上市 行 业 7 月 2 日，特斯拉公布其今年第二季度的全球销量达 384122 辆， 较去年同期的 443956 辆下降 13%，已连续第二个季度出现下滑。这一 下滑主要归因于公司更实惠车型上市延迟、首席执行官埃隆·马斯克 的政治立场争议以及在美国市场销量的大幅下滑。 定 期 报 告 特斯拉目前面临着多重困境交织的局面。美国国会通过巨额减税 与支出法案，其中取消电动车税收抵免政策的部分，将对特斯拉构成 潜在冲击。同时面对通用汽车、大众、宝马、比亚迪、小米等推出的 新车型，特斯拉沿用多年的产品线正在承受越来越大的竞争压力。而 面对汽车业务的疲软，特斯拉积极挖掘自动驾驶软件、无人驾驶网约 车服务和人形机器人 Optimus 的发展潜力。 7 月 3 日，小鹏汽车正式发布全新 AI 智能家庭 SUV——小鹏 G7。 新车共推出 602 长续航 Max、702 超长续航 Max、702 超长续航 Ultra 三个版本，官方指导售价分别为 19.58 万元、20.58 万元和 22. ...

Core Viewpoint - Huawei's announcement of the open-source Pangu Pro MoE model with 720 billion parameters and the Pangu model with 70 billion parameters has sparked industry discussions regarding the similarities between the Pangu Pro MoE model and Alibaba's Tongyi Qwen-2.5 14B model in terms of parameter structure [1][2]. Group 1 - The Noah's Ark Lab stated that the Pangu Pro MoE model is developed and trained based on the Ascend hardware platform and is not an incremental training of other vendors' models [1]. - A GitHub study found a high correlation of 0.927 in attention parameter distribution between the Pangu Pro MoE model and the Tongyi Qwen-2.5 model, indicating significant structural similarities [1]. - Noah's Ark Lab clarified that while some code implementations of the Pangu Pro MoE model reference industry open-source practices, they strictly adhere to open-source license requirements and clearly mark copyright statements [1][2]. Group 2 - Industry analysts suggest that the Pangu Pro MoE model likely did not use the pre-trained weights of the Tongyi Qwen-2.5 model as initialization parameters, as there are essential differences in the absolute value distribution of biases [2]. - The structural consistency between the two models may stem from shared architectural design principles, which is common in large models as good structures are widely adopted [2]. - Noah's Ark Lab emphasized that the Pangu Pro MoE model features key innovations, being the first mixed expert model designed for the Ascend hardware platform, and introduced a grouped mixed expert model (MoGE) architecture to enhance training efficiency [2]. Group 3 - Noah's Ark Lab expressed gratitude to global developers and partners for their support of the Pangu model and highlighted the importance of constructive feedback from the open-source community [3]. - The lab aims to optimize model capabilities through collaboration with like-minded partners, accelerating technological breakthroughs and industry applications [3].

Core Viewpoint - Longqi Technology, a leading ODM manufacturer in smart hardware, has submitted its listing application to the Hong Kong Stock Exchange after just over a year on the Shanghai Stock Exchange, aiming to leverage its strong market position and diverse product offerings in the consumer electronics sector [1][2]. Group 1: Business Model and Market Position - Longqi Technology provides comprehensive ODM services, integrating hardware design, software development, testing, and supply chain management, with a diverse product range including smartphones, AI PCs, automotive electronics, tablets, smartwatches, and smart glasses [1]. - The company is the second-largest consumer electronics ODM globally and ranks first in smartphone ODM shipments as of 2024 [1]. - Major clients include leading brands such as Xiaomi, Samsung, Lenovo, Honor, OPPO, and vivo [1]. Group 2: Financial Performance - Revenue figures for Longqi Technology from 2022 to 2024 are as follows: 29.343 billion RMB, 27.185 billion RMB, and 46.382 billion RMB, while net profits were 562 million RMB, 603 million RMB, and 493 million RMB respectively [1][4]. - The company's gross profit increased over the same period, but the gross margin showed a decline, with figures of 8.1%, 9.5%, and 5.8% for the respective years [4][5]. - In Q1 2025, the company reported a revenue of 9.378 billion RMB, a year-on-year decrease of 9.27%, but net profit increased by 20.33% to 154 million RMB, indicating a recovery in profitability [4]. Group 3: Strategic Growth Initiatives - Longqi Technology has adopted a "1+2+X" strategy to align with new market demands and technological innovations, focusing on smartphones as the core, with personal computing and automotive electronics as key growth areas, and expanding into emerging consumer electronics [2]. - The AIoT and other product segments have shown continuous revenue growth, with figures of 1.887 billion RMB, 2.511 billion RMB, and 5.573 billion RMB from 2022 to 2024, increasing their revenue share from 6.5% to 12% [3][4]. Group 4: Industry Trends and Future Outlook - The rise of AI technology is expected to create significant growth opportunities for Longqi Technology, particularly in the smartphone and PC markets, where AI penetration rates are projected to reach 18% and 32% respectively in 2024 [6]. - Longqi Technology has been proactive in developing AIoT products, launching over 10 million smartwatches in 2023 and achieving significant sales in smart glasses and VR products since 2015 [6][7]. - The company is well-positioned to benefit from the anticipated growth in AI hardware products, which could enhance its performance and appeal to growth-oriented investors in the Hong Kong capital market [8].

Core Viewpoint - The RuiPath pathology model, developed by Ruijin Hospital in collaboration with Huawei, aims to enhance the efficiency and accuracy of pathology diagnostics in China by leveraging AI technology [1][5]. Group 1: Model Development and Features - The RuiPath model is a clinical-grade multimodal pathology model that covers 90% of the annual cancer incidence in China, addressing 19 common cancer types and hundreds of auxiliary diagnostic tasks [1][5]. - The model has achieved state-of-the-art (SOTA) performance in 7 out of 14 auxiliary diagnostic tasks tested against 12 mainstream public datasets, surpassing the performance of Harvard's UNI2 model [4]. - The model's core "visual foundation model" was developed using over one million high-quality digital pathology slides from Ruijin Hospital, utilizing Huawei's AI toolchain for annotation, training, and fine-tuning [2][4]. Group 2: Efficiency and Impact - The implementation of the RuiPath model allows pathologists to increase their daily workload from 200-300 slides to 400-500 or more, significantly improving diagnostic efficiency [5]. - The model aims to standardize digital pathology practices across hospitals in China, enabling easier deployment and reducing training costs for other institutions [5][10]. - The collaboration between Ruijin Hospital and Huawei has streamlined the model training process, allowing for the completion of the RuiPath model development with only a 16-card cluster, making it more accessible for hospitals [10][11]. Group 3: Industry Challenges and Solutions - There is a significant shortage of pathology doctors in China, with only about 20,000 available and a gap of 140,000 needed, highlighting the importance of AI solutions in addressing this challenge [5]. - The partnership has evolved through two phases: digitalization and smart pathology, focusing on data standardization and collaborative model development [7][8]. - The use of Huawei's ModelEngine has transformed the annotation process, allowing pathologists to review over 700 slides in a day, thus enhancing both efficiency and accuracy [10].

注意看，这张图画的是谁？ 如果不够明显，那再眯着眼睛看呢。没错，就是蒙娜丽莎。 这张来自日本艺术家北冈秋吉前几天创作的一张图，结果直接把一众大模型难倒了。 ChatGPT只能分辨出这是一张脸。。 Gemini则是直接识别错了人。 白交 发自 凹非寺 量子位 | 公众号 QbitAI 又一个让大模型几乎全军覆没的难题出现了。 大模型不会眯眼睛？ 为了验证这一推论，咱们也进行了一波简单实测。既然直接问「画的是谁」问不出来，那给一点提示，让它可以尝试眯着眼睛。 网友：懂了，大模型不会眯眼睛。 首先来看ChatGPT的表现。 它能准确判断出这是一张视觉错位图，也通过「眯着眼睛」识别出来了具体的脸部轮廓，结果在最后关头识别错了。 再来看看国产玩家的表现，还是挑出三个来看看。 豆包的回答与Gemini类似，能够判断出图像的风格、人脸轮廓，但是无法判断具体任务。 但选择 深度思考 之后，它深入研究了故障艺术的通道分离特性，然后从黑色轮廓中判断出这是 爱因斯坦 （？？？？ 而再让它深度思考一下，答案给不出来了。 Gemini的回答则还停留在各种颜色的条纹，也辨认出这是个侧脸的轮廓。 Grok则直接表示无法辨认，让我提供一张更 ...