风险提示：政策变动、行业竞争加剧、消费环境修复不及预期等。 投资建议：维持"优于大市"评级。伴随国家不断出台扩大内需的利好政策，报告期内社服板块估值有望 持续修复。目前经济环境和市场风格下建议配置携程集团-S、亚朵、同程旅行、BOSS 直聘、小菜园、 古茗、蜜雪集团、卓越教育集团、长白山等。中线优选中国中免、美团-W、蜜雪集团、古茗、亚朵、 华住集团-S、卓越教育集团、学大教育、立国际控股、中国东方教育、携程集团-S、海底捞、同程旅 行、北京人力、BOSS 直聘、科锐国际、锦江酒店、同庆楼、宋城演艺、百胜中国、九毛九、海南机 场、君亭酒店、峨眉山A、天目湖、三特索道、黄山旅游、广州酒家、首旅酒店、东方甄选、曹操出行 等。 周观点：凯文教育携手智谱华章，成立合资公司布局AI 教育。凯文教育是A股稀缺的K12 学校运营服务 商，据公司年报，业务涵盖三大板块：K12 学校业务、素质教育业务（提供校内外体育、艺术等特色培 训）、职业教育服务（产教融合及党校业务）。2025 年上半年，公司实现营业收入1.71 亿元，同比增 长12.51%，归母净利润实现扭亏为盈。近期发布公告与智谱华章成立合资公司，业务版图进一步扩展 ...

Core Insights - The article discusses the evolution of the AI landscape in China, highlighting the shift from a competitive "hundred models war" to a focus on application ecosystems, characterized by the emergence of "invisible sects" linked by technology, talent networks, and capital [1] - It contrasts two main factions: the "Tsinghua system," represented by companies like Zhipu and Moonlight, and the "Alibaba system," represented by entrepreneurs from Alibaba, both of which are shaping the future of the domestic AI industry [1] Origin: Academic Roots and Industrial Foundations - The "Tsinghua system" traces its origins to the Knowledge Engineering Group (KEG) at Tsinghua University, led by Professor Tang Jie, focusing on knowledge graphs, graph neural networks, and pre-trained models, embodying a traditional academic research approach [1][3] - Zhipu, as a direct descendant of KEG, aims to commercialize decades of research, led by CEO Zhang Peng, who emphasizes a theoretical-driven path distinct from mainstream models like GPT and BERT [3] - Moonlight, founded by Yang Zhilin, combines theoretical depth with engineering execution, leveraging international experience to create innovative products like the Kimi intelligent assistant, which supports extensive context input [5] Divergence: Technical Lineage and Entrepreneurial Orientation - The "Tsinghua system" is characterized by a "theory-driven innovation" approach, focusing on fundamental model architecture challenges, as seen in Zhipu's GLM series and Moonlight's emphasis on long-text processing capabilities [10][12] - In contrast, the "Alibaba system" adopts a "scene-driven engineering" approach, optimizing model deployment around specific business needs, emphasizing product efficiency and industry adaptability [12] - The founders of the "Tsinghua system" often come from academic backgrounds, while the "Alibaba system" features battle-hardened entrepreneurs with a pragmatic, market-sensitive approach [12][13] Competition and Cooperation: Complex Relationships - The competition between the "Tsinghua system" and "Alibaba system" revolves around attracting top AI talent, GPU resources, and defining the next generation of AI applications, with both sides vying for market leadership [14] - Despite their rivalry, there are cooperative elements, as Alibaba strategically invests in promising startups from the "Tsinghua system," creating a complex "co-opetition" dynamic [14][16] - This relationship allows Alibaba to maintain its technological edge while also integrating cutting-edge innovations from external startups into its ecosystem [16] Future Directions: Defining New Paradigms - The rise of both systems reflects the diversity of AI development paths in China, emphasizing the need for integration between theoretical depth and commercial acumen [17] - Future competition will hinge on the ability of both factions to adapt, with "Tsinghua system" researchers needing to transition into product-oriented roles, while "Alibaba system" entrepreneurs must deepen their technical foundations [17] - The ultimate outcome may not be a single dominant faction but the emergence of new AI enterprises that blend the strengths of both systems, fostering a more mature and competitive landscape [17]

Core Insights - The GLM-4.5 model focuses on expanding context length and improving its hallucination prevention capabilities through effective Reinforcement Learning from Human Feedback (RLHF) processes [6][10][11] - The future development will prioritize reasoning, programming, and agent capabilities, with plans to release smaller parameter models [6][50][28] Group 1: GLM-4.5 Development - The team behind GLM-4.5 includes key contributors who have worked on various significant AI projects, establishing a strong foundation for the model's development [3] - The choice of GQA over MLA in the architecture was made for performance considerations, with specific weight initialization techniques applied [12][6] - There is an ongoing effort to enhance the model's context length, with potential releases of smaller dense or mixture of experts (MoE) models in the future [9][28] Group 2: Model Performance and Features - GLM-4.5 has demonstrated superior performance in tasks that do not require long text generation compared to other models like Qwen 3 and Gemini 2.5 [9] - The model's effective RLHF process is credited for its strong performance in preventing hallucinations [11] - The team is exploring the integration of reasoning models and believes that both reasoning and non-reasoning models will coexist and complement each other in the long run [16][17] Group 3: Future Directions and Innovations - The company plans to focus on developing smaller MoE models and enhancing the capabilities of existing models to handle more complex tasks [28][50] - There is an emphasis on improving data engineering and the quality of training data, which is crucial for model performance [32][35] - The team is also considering the development of multimodal models, although current resources are primarily focused on text and vision [23][22] Group 4: Open Source vs. Closed Source Models - The company believes that open-source models are closing the performance gap with closed-source models, driven by advancements in resources and data availability [36][53] - The team acknowledges that while open-source models have made significant strides, they still face challenges in terms of computational and data resources compared to leading commercial models [36][53] Group 5: Technical Challenges and Solutions - The team is exploring various technical aspects, including efficient attention mechanisms and the potential for integrating image generation capabilities into language models [40][24] - There is a recognition of the importance of fine-tuning and optimizing the model's writing capabilities through improved tokenization and data processing techniques [42][41]

Core Insights - The domestic open-source large model sector is experiencing significant growth, drawing widespread attention from the industry [1] - A notable observation is that the top-ranking open-source AI models on the Design Arena platform are predominantly from China [1][2] Group 1: Model Rankings - The Design Arena platform employs a unique evaluation mechanism where users vote on responses generated by different models, ensuring fairness and dynamism in rankings [2] - Among the top 15 models listed as open-source, all positions are occupied by Chinese models, with DeepSeek-R1-0528 leading the list, followed by Zhizhu's GLM-4.5 and Alibaba's Qwen 3 Coder 480B [2][3] - The ranking showcases multiple models from various manufacturers, including DeepSeek, Qwen, and GLM, with the first non-Chinese model, OpenAI's GPT OSS 120B, appearing only at the 16th position [2][3] Group 2: Industry Developments - Recent releases of new-generation open-source large models by domestic AI companies are propelling advancements in AI technology [4] - A total of 33 large models from various manufacturers, including Alibaba and Zhizhu, were released in July, indicating a robust trend in the domestic open-source model landscape [4] - The emergence of 19 leading open-source model laboratories in China, such as DeepSeek and Qwen, highlights the collaborative efforts driving the rise of domestic open-source models [4] Group 3: Competitive Landscape - Historically, closed-source models like the GPT series have maintained a technological edge, but the rise of open-source models, particularly the Llama series, is reshaping the global AI landscape [4] - Chinese open-source models like Qwen and DeepSeek are now recognized as competitive alternatives to top-tier closed-source models, facilitating a shift in focus towards model tuning and application optimization in the industry [4]

Core Viewpoint - Haidian District is emerging as a significant hub for AI innovation, attracting talent and investment, and fostering a robust ecosystem that supports the development of AI technologies and applications [2][3][4]. Group 1: Haidian's Innovation Ecosystem - Haidian has become a focal point for tech innovation, with over 20,000 external investment personnel active monthly and numerous unicorns emerging from the area [2]. - The district accounts for 2.6% of Beijing's land but generates over 25% of the city's GDP, hosting more than 70% of the nation's AI companies and 80% of top global AI scholars [3]. - The area is home to over 100 AI companies, establishing itself as the core of the "Zhongguancun AI Large Model Industry Cluster" [3]. Group 2: Historical Context and Development - Haidian has historically been linked to every wave of AI development in China, from early expert systems to the current era of deep learning and large models [6][7]. - The establishment of key research institutions and collaborations with leading universities has laid a strong foundation for AI research and talent cultivation [9][10]. Group 3: AI Application and Market Potential - The AI application market is viewed as a trillion-dollar opportunity, with Haidian at the center of this entrepreneurial resurgence [4][5]. - The district has seen a resurgence in startup activity, reminiscent of the mobile internet boom, with numerous events and networking opportunities for entrepreneurs [4]. Group 4: Infrastructure and Support Mechanisms - Haidian is implementing a comprehensive strategy to support AI development, including a public computing power platform and a data-sharing initiative [12][13]. - The district has established a significant number of large models, with 89 registered by June 2023, representing one-third of the national total [13]. Group 5: Talent and Investment - Haidian boasts the highest concentration of AI talent in China, with 80% of the nation's top AI scholars and numerous educational institutions offering AI programs [14]. - The district has launched a series of funds totaling 20 billion yuan to support technology companies throughout their growth cycles, enhancing its investment landscape [14][15].

Group 1 - The core players in China's foundational model landscape are ByteDance, Alibaba, Jiyue Xingchen, Zhipu AI, and DeepSeek, collectively referred to as the "Five Strong" [1] - DeepSeek is recognized as a strong technical dark horse due to its breakthroughs in mathematical reasoning and cost-effectiveness, while ByteDance holds a comprehensive advantage with its full-stack layout and extensive user ecosystem [13][25] - Alibaba maintains its position as the king of open-source models, leveraging top-tier global open-source models and infrastructure, although it faces challenges in deepening commercialization [13][25] Group 2 - Jiyue Xingchen is noted for its multi-modal technology and rapid rise in terminal applications, but it needs to address the challenge of achieving an integrated architecture [11][25] - Zhipu AI, while having a solid presence in the government and enterprise market, is limited by its reliance on traditional technology paths and has not demonstrated disruptive breakthroughs [12][25] - The future competitive landscape will focus on three dimensions: DeepSeek's reasoning capabilities, how ByteDance and Alibaba convert their ecosystems into commercial success, and whether Jiyue Xingchen can overcome multi-modal integration challenges [16][23] Group 3 - DeepSeek excels in specialized fields like mathematical reasoning but has a relatively narrow commercial application scope, which may put it at a disadvantage in overall competition [22][25] - Zhipu AI's strong academic background is countered by its limited consumer applications and over-reliance on the B-end market, which weakens its risk resistance [22][25] - In contrast, Alibaba, ByteDance, and Jiyue Xingchen demonstrate stronger overall capabilities with tighter integration of technology and business [22][25] Group 4 - The competitive key points include the intelligence ceiling defined by model reasoning capabilities, the importance of multi-modal capabilities as a foundation for AGI, and the need for continuous validation of market acceptance for open-source ecosystems and vertical applications [23][25] - Alibaba and ByteDance are currently leading the first tier due to their comprehensive funding, ecosystem, and technology layouts, while Jiyue Xingchen shows significant potential with its multi-modal technology [23][25] - DeepSeek and Zhipu AI need to continue making breakthroughs in differentiated areas to remain competitive [23][25]

坚持最高安全标准和公平、透明、可持续原则，推动人工智能技术发展。 智谱强调，其在技术上坚定投入。基于自主原创的GLM系列模型，以Model as a Service(MaaS) 为理念构建的开放平台 bigmodel.cn，正持续为千行百业带来人工智能创新变革。C端产品智谱清 言，帮助数千万中国用户使用国产AI助手，用AI服务社会。 " 鉴于智谱掌握全链路大模型核心技术的事实，被列入实体清单不会对公司业务产生实质影响。智谱 有能力也将更专注地为我们的用户和伙伴提供世界一流的大模型技术、产品和服务。同时公司将继续 参与全球人工智能竞争，坚持最高安全标准和公平、透明、可持续原则，推动人工智能技术发展。 "智谱如是表示。 本文由公众号IPO早知道（ID：ipozaozhidao）原创撰写，如需转载请联系C叔↓↓↓ 本文为IPO早知道原创 作者｜Stone Jin 微信公众号｜ipozaozhidao 据IPO早知道消息，美国商务部工业和安全局（BIS）于北京时间2025年1月15日晚间新增加一批实 体列入实体清单，其中包括中国头部大模型企业智谱旗下多个实体。 对此， 智谱第一时间发布声明称， 这一决定缺乏事实依据， ...