2025年08月25日 证券研究报告 | 人工智能行业专题（11） 探究模型能力与应用的进展和边界 行业研究 · 行业专题 互联网 · 互联网II 投资评级：优于大市（维持） 证券分析师：张伦可 证券分析师：陈淑媛 证券分析师：刘子谭 证券分析师：张昊晨 0755-81982651 021-60375431 liuzitan@guosen.com.cn zhanghaochen1@guosen.com.cn zhanglunke@guosen.com.cn chenshuyuan@guosen.com.cn S0980525060001 S0980525010001 S0980521120004 S0980524030003 请务必阅读正文之后的免责声明及其项下所有内容 报告摘要 Ø 风险提示：宏观经济波动风险、广告增长不及预期风险、行业竞争加剧风险、AI技术进展不及预期风险等。 请务必阅读正文之后的免责声明及其项下所有内容 2 Ø 本篇报告主要针对海内外模型发展、探究模型能力与应用的进展和边界。我们认为当前海外模型呈现差异化发展，企业调用考虑性价比。当前 OpenAI在技术路径上相对领先，聚焦强化推理与专业 ...

Core Viewpoint - DeepSeek's recent update to version V3.1 has disappointed many users who were eagerly awaiting the release of the R2 model, which has been delayed for several months, leading to a decline in the company's popularity and user engagement [2][3][10] Group 1: Product Updates - DeepSeek released V3.1 on August 21, which ranked third on HuggingFace's trend list, but many users expressed dissatisfaction and called for the return of the previous R1 model [2][3] - The V3.1 update features a hybrid reasoning architecture that combines thinking and non-thinking modes, enhancing efficiency and aligning with trends seen in other major models like GPT-5 [4] - V3.1 offers faster response times and improved agent capabilities, with an expanded context of 128K after the API upgrade [5] Group 2: Pricing Changes - Starting September 6, DeepSeek will adjust its API pricing to 0.5 RMB per million tokens for cache hits, 4 RMB for cache misses, and 12 RMB for output, representing a middle ground between previous versions [5] Group 3: Competitive Landscape - Other domestic AI models, such as those from Zhiyu and Alibaba, are rapidly updating and releasing new features, creating a competitive environment that DeepSeek is struggling to keep up with [7][8] - The overall market for large models is intensifying, with significant advancements from both domestic and international competitors, including OpenAI's GPT-5 and Google's Genie 3 [9] Group 4: User Engagement and Market Position - DeepSeek's website traffic has been declining for four consecutive months, with a 9.63% average monthly decrease, and its app's monthly active users fell to 82.93 million in July, marking a significant drop [10]

Core Viewpoint - The recent upgrade of DeepSeek to version 3.1 has shown significant improvements in context length and user interaction, while also merging features from previous models to reduce deployment costs [1][11][12]. Group 1: Model Improvements - DeepSeek V3.1 now supports a context length of 128K, enhancing its ability to handle longer texts [4]. - The model's parameter count increased slightly from 671 billion to 685 billion, but the user experience has improved noticeably [5]. - The model's programming capabilities have been highlighted, achieving a score of 71.6% in multi-language programming tests, outperforming Claude 4 Opus [7]. Group 2: Economic Efficiency - The merger of V3 and R1 models allows for reduced deployment costs, requiring only 60 GPUs instead of the previous 120 [12]. - Developers noted that the performance could improve by 3-4 times with the new model due to increased cache size [12]. - The open-source release of DeepSeek V3.1-Base on Huggingface indicates a move towards greater accessibility and collaboration in the AI community [13]. Group 3: Market Context - The AI industry is closely watching the developments of DeepSeek, especially in light of the absence of the anticipated R2 model [19]. - Competitors like OpenAI, Google, and Alibaba have released new models, using R1 as a benchmark for their advancements [1][15]. - The market is eager for DeepSeek's next steps, particularly regarding the potential release of a multi-modal model following the V3.1 update [23].

Core Insights - DeepSeek's recent announcement of a $700 million funding round was quickly retracted, leading to confusion and speculation within the investment community [1][3] - Despite the funding rumors, DeepSeek has not publicly disclosed previous funding rounds and appears to be in a strong financial position, with significant backing from state-owned entities and a large budget for research [3][4] - The company faces challenges with its upcoming R2 model, which has been delayed and is under scrutiny for not outperforming its predecessor, R1, in key performance metrics [4][6] Financial Position - DeepSeek reportedly incurs substantial operational costs, including $700 million annually for server expenses and high salaries for talent acquisition [6] - The management scale of its partner, Huanfang Quantitative, has decreased from a peak of $100 billion in 2021 to $45 billion, indicating a significant contraction in the investment landscape [6] - The company is under pressure to secure additional funding as its financial burn rate accelerates, prompting recruitment for key financial positions [6] Market Dynamics - The competitive landscape is intensifying, with major players like OpenAI and Google launching new products that overshadow DeepSeek's silence and delays [6][8] - There is a growing concern among investors regarding DeepSeek's ability to deliver on its promises of low-cost, high-performance technology, which could shift perceptions from "technological idealism" to "inadequate capabilities" [6][8] - The anticipation surrounding the release of R2 is critical, as it must meet high performance standards and competitive pricing to maintain investor confidence and market position [8]

Group 1 - The core viewpoint highlights the strong performance of the A-share "chip giant" Cambricon, with its stock price rising over 14% to reach 985 yuan, resulting in a market capitalization exceeding 395 billion yuan [1] - The Huabao Science and Technology Artificial Intelligence ETF (589520), which has Cambricon as its top holding, saw significant trading activity, with a peak price increase of 1.5% before a slight decline [1] - The ETF has experienced a premium in the market, indicating strong buying interest, and has attracted a total of 821.7 million yuan in investments over the past 60 days, reflecting optimism towards the AI sector [1] Group 2 - The competitive landscape for large models is intensifying, with the latest rankings showing that nine out of the top ten models are domestic, led by Zhipu's GLM-4.5 [3] - Despite DeepSeek R1 not ranking highly, it has received the most likes among domestic models, indicating strong user engagement [3] - Analysts suggest that the trend towards self-sufficient computing chips is inevitable, with domestic cloud providers likely to adapt to local chip solutions due to security concerns with Nvidia chips [3] Group 3 - Historical trends indicate that sectors that have undergone early adjustments can initiate a second wave of growth, with AI meeting conditions for renewed leadership in the market [4] - The AI sector has seen a period of adjustment followed by gradual recovery, suggesting a potential for further upward movement [4] Group 4 - The Huabao Science and Technology Artificial Intelligence ETF (589520) is positioned as a key player in the domestic AI industry, with a significant focus on semiconductor stocks, which account for nearly half of its top holdings [5] - The ETF is designed to benefit from the integration of edge and cloud computing, with a balanced allocation across software applications, terminal applications, and chips [5]

Core Insights - OpenAI has officially launched GPT-5, which is described as the most intelligent, fastest, and useful model to date by CEO Sam Altman [1][2] Model Highlights - GPT-5 is a fusion model that automatically adjusts its thinking depth based on the complexity of the question [2][7] - It has achieved record high scores in various industry benchmarks, including 94.6% accuracy in the AIME 2025 math test, 84.2% in multi-modal understanding, and 46.2% in the HealthBench Hard medical test [4] - The model significantly reduces the "hallucination" problem and is more honest about its capabilities [2][7] Programming Capabilities - GPT-5 shows remarkable improvements in programming, scoring 74.9% in the SWE-bench Verified test and 88% in the Aider polyglot test [4] - It can generate complex code quickly, as demonstrated by creating a complete French learning game in seconds [4] Medical Applications - GPT-5 is touted as the most accurate model for medical queries, enhancing patient understanding and decision-making [6] - It is designed to complement, not replace, doctors by improving patient knowledge and communication [6] Commercialization Strategy - OpenAI has raised $8.3 billion, with a valuation of $300 billion, and its annual recurring revenue has increased from $10 billion to $13 billion [8] - The launch of GPT-5 comes amid intense global AI competition, with other companies like Google and Meta also advancing their models [8] Market Positioning - OpenAI is actively expanding into enterprise and government markets, offering ChatGPT enterprise versions at a symbolic price to federal agencies [8][9] - The company has signed a $200 million contract with the U.S. Department of Defense to explore AI applications in various fields [9] Competitive Landscape - In the enterprise AI market, OpenAI holds a 25% share, trailing behind Anthropic (32%) and Google (20%) [10] - The ability of GPT-5 to solve complex problems may create differentiated economic value in high-margin sectors like strategic consulting and investment analysis [10]

Core Insights - The AI chess tournament hosted on Kaggle featured eight large language models (LLMs) competing in a knockout format, with Grok 4 and o3 advancing to the finals after defeating Gemini 2.5 Pro and o4-mini respectively [1][3][8] Group 1: Tournament Structure and Results - The tournament lasted three days and involved eight AI models, including Grok 4 (xAI), Gemini 2.5 Pro (Google), o4-mini (OpenAI), o3 (OpenAI), Claude 4 Opus (Anthropic), Gemini 2.5 Flash (Google), DeepSeek R1 (DeepSeek), and Kimi k2 (Moonshot AI) [1] - The competition utilized a single-elimination format where each AI had up to four attempts to make a legal move; failure to do so resulted in an immediate loss [1] - On the first day, Grok 4, o3, Gemini 2.5 Pro, and o4-mini all achieved 4-0 victories, advancing to the semifinals [3][11][22] Group 2: Semifinal Highlights - In the semifinals, o3 demonstrated a dominant performance, winning 4-0 against o4-mini, showcasing a high level of precision with a perfect accuracy score of 100 in one of the games [5] - The match between Grok 4 and Gemini 2.5 Pro ended in a tie after regular play, leading to an Armageddon tiebreaker where Grok 4 emerged victorious [8] - The semifinals highlighted the strengths and weaknesses of the AI models, with Grok 4 overcoming early mistakes to secure its place in the finals [8][19] Group 3: Performance Analysis - The tournament revealed that while some AI models performed exceptionally well, others struggled with basic tactical sequences and context understanding, indicating areas for improvement in AI chess capabilities [22] - The performance of Grok 4 attracted attention from industry figures, including Elon Musk, who commented on its impressive gameplay [19]

Core Viewpoint - OpenAI is re-entering the open-source arena by launching two significant models, gpt-oss-120b and gpt-oss-20b, amidst a growing competition between open-source and closed-source AI solutions [1][2]. Group 1: OpenAI's Historical Context - Since its establishment in 2015, OpenAI has been at the forefront of AI technology, with the launch of ChatGPT in 2022 marking a significant milestone in user growth [1]. - Initially, OpenAI embraced open-source principles with the releases of GPT-1 and GPT-2, but shifted to a closed-source model with GPT-3 in 2020, which drew criticism for contradicting its mission to benefit humanity [1][2]. Group 2: New Open-Source Models - The newly launched models, gpt-oss-120b and gpt-oss-20b, are designed for high inference in cloud environments and low latency on edge devices, providing developers with more options [2]. - The release of these models has generated significant interest in the AI open-source community, leading to a surge in downloads on the Hugging Face platform, prompting requests to manage server load [2]. Group 3: Industry Reactions and Implications - Opinions on OpenAI's approach are mixed; some view it as a protective measure for core assets, while others argue it limits developers' ability to conduct in-depth research and hinders the open-source ecosystem [4]. - OpenAI's collaboration with cloud providers like Amazon AWS is seen as a strategy to enhance the distribution and application of its open-source models [4]. - The performance gap between open-source and closed-source models is narrowing, increasing competition from global open-source players, particularly in the Chinese market with Alibaba's Qwen series achieving over 300 models and 400 million downloads by July 2023 [4]. Group 4: Future Outlook - OpenAI's re-entry into open-source presents both challenges and opportunities, suggesting that AI giants may adopt more flexible strategies in response to the rapidly changing market [7].

Core Viewpoint - The article discusses the first Kaggle AI chess competition initiated by Google, highlighting the performance of various AI models, particularly Grok 4, which has shown exceptional capabilities in tactical strategy and speed during the matches [2][16]. Group 1: Competition Overview - The Kaggle AI chess competition is designed to promote the Kaggle gaming arena, with chess as the inaugural event [6]. - The competition features AI models from OpenAI, DeepSeek, Kimi, Gemini, Claude, and Grok [7]. - Matches are being live-streamed daily from August 5 to August 7, starting at 10:30 AM Pacific Time [8]. Group 2: Performance Highlights - Grok 4 emerged as the best performer in the initial round, while DeepSeek R1 showed strong performance but lost to o4-mini [2][12]. - The quarterfinals saw Grok 4 and Gemini 2.5 Pro advance, alongside ChatGPT's o4-mini and o3 [12]. - Grok 4's performance was likened to that of a "real GM," showcasing its tactical prowess [17]. Group 3: Match Analysis - In the match between Grok 4 and Gemini 2.5 Flash, Grok 4 dominated, while Gemini Flash struggled from the start [18]. - The match between OpenAI's o4-mini and DeepSeek R1 highlighted R1's initial strong opening but ultimately led to its defeat due to critical errors [20][21]. - The best match of the day was between Gemini 2.5 Pro and Claude Opus 4, where both models displayed high-level chess skills, although Claude made some mistakes [23]. Group 4: AI Evaluation - The competition serves as a test of AI's emergent capabilities, with chess being an ideal scenario due to its complex yet clear rules [31][36]. - The article notes that AI's strength in this context comes from its ability to generalize rather than from task-specific training [38]. - There is a general consensus among observers that chess is a reliable method for assessing AI capabilities [39]. Group 5: Public Sentiment and Predictions - Prior to the competition, Gemini 2.5 Pro was favored to win, but Grok 4 gained overwhelming support after the quarterfinals [42][44]. - The article humorously speculates on future AI competitions, suggesting games like UNO could be next [40].

Core Viewpoint - The article discusses the results of the first large model chess competition organized by Google, highlighting the performance of various AI models, particularly Grok 4, which emerged as a strong contender with a perfect record [2][30]. Group 1: Competition Overview - The chess competition lasted three days and featured models such as Gemini 2.5 Pro, o4-mini, Grok 4, and o3, all achieving a 4-0 victory in the first round [4]. - The competition was held on the Kaggle Game Arena platform, aiming to evaluate the performance of large language models (LLMs) in dynamic and competitive environments [6]. Group 2: Match Results - Kimi k2 lost to o3 with a score of 0-4, failing to make legal moves in all four games [7][8]. - o4-mini defeated DeepSeek R1 with a score of 4-0, showcasing a decline in game quality after a few strong opening moves [18][21]. - Gemini 2.5 Pro won against Claude 4 Opus with a score of 4-0, although its true strength remains uncertain due to Claude's mistakes [23][24]. - Grok 4 achieved a perfect score of 4-0 against Gemini 2.5 Flash, demonstrating superior chess skills and the ability to capitalize on unprotected pieces [30][33]. Group 3: Key Observations - The competition revealed three main weaknesses in current AI models: insufficient global board visualization, limited understanding of piece interactions, and issues executing legal moves [36]. - Grok 4's performance suggests it may have overcome these limitations, raising questions about the stability of these advantages in future matches [36]. Group 4: Audience Engagement - A poll conducted prior to the competition indicated that 37% of participants favored Gemini 2.5 Pro as the likely winner, with Grok 4 receiving 7.04% of the votes [37][38].