专题：亏损超230亿！美团公布2025账单 炒股就看金麒麟分析师研报，权威，专业，及时，全面，助您挖掘潜力主题机会！ 来源：道术意义 今年春节，50亿元的营销费用砸向了AI导流领域。 阿里+通义千问，腾讯+元宝，字节跳动+豆包，AI三国杀，三大巨头都到齐了。 Tier1级别的对手厮杀，每天上千万的"AI点奶茶"营销，AI导流截胡令传统营销路子相形见绌。 Tier2的头马美团似乎支撑不住。 2月13日，如市场所预料，美团公布了预亏损报告，预计2025年全财年录得亏损233-243亿元，同比去年 暴跌约600亿元之巨。 分基地与主基地 确实凶险。 按照电竞《星际争霸》的打法，当对手开了分基地，用分基地兵力攻击我主基地时候，战局渐渐倾向于 一边倒。 因为对手牺牲了分基地的发展，不断拖耗你主基地的资源，而自己的主基地不受攻击，持续成长。 按照数学演算，如果没有打乱战局的妙策，我主基地倒下只是一个攻击力数值与时间的函数。 2025年一年，这个戏剧性的情况是怎么造成的？ 我们来看几个重要的时间点和抽样数据统计。 第一个时间点是2025年6月，阿里蒋凡回归。 从2021年12月被"流放"海外，到2024年11月重回权力中心 ...

Core Insights - The article discusses the intense competition among internet giants in the AI large model sector, highlighting the ambitions of companies to establish their AI applications as the primary traffic entry point [4][23] - DeepSeek, founded by Liang Wenfeng, emerged as a significant player in the AI landscape with its R1 model, which was launched at a surprisingly low cost, challenging the perception of high investment requirements for top-tier models [14][31] - Despite the competitive environment, DeepSeek has maintained a low profile, with recent updates suggesting a potential new model release, V4, but with no official confirmation [26][27] Industry Competition - Major companies are aggressively distributing cash incentives to attract users, with Tencent offering 1 billion yuan, Baidu 500 million yuan, and Alibaba 3 billion yuan, indicating a fierce battle for user engagement [25] - The launch of new models by ByteDance and Alibaba, including the 2.0 versions of their respective models, reflects a rapid evolution in AI capabilities and competition [8][25] - The article notes a peculiar competitive dynamic where companies are responding to each other's moves, creating a sense of mutual awareness in the market [8][25] DeepSeek's Position - DeepSeek's recent updates include an increase in context window length from 128K tokens to 1 million tokens, suggesting advancements in their technology [26] - The company continues to recruit talent despite a slowdown in hiring across the industry, indicating its commitment to innovation and development [27] - Liang Wenfeng's vision for DeepSeek is to lead in AI research and development, aiming to create a general-purpose AI that goes beyond existing models [31][32] User Engagement and Market Dynamics - The article emphasizes the importance of addressing user needs in the AI sector, with companies like DeepSeek beginning to focus on consumer-facing products [33] - The competition is framed as a quest to meet real user demands, which will determine the leading players in the AI landscape [36] - The article concludes that the current battle among internet giants is crucial for defining the next decade of internet order, highlighting the strategic significance of user engagement in AI applications [36]

Core Viewpoint - The article discusses the intense competition among major internet companies in China to dominate the AI model application space, highlighting the strategic positioning of Deep Seek and its founder Liang Wenfeng as a significant player in this evolving landscape [2][4]. Group 1: AI Competition Landscape - Major internet giants are aggressively investing in user incentives, with Tencent distributing 1 billion yuan in cash red envelopes, Baidu offering 500 million yuan for promoting its Wenxin assistant, and Alibaba launching a 3 billion yuan campaign [4]. - The competition is characterized by rapid product releases, with ByteDance announcing its Doubao model 2.0 and Alibaba introducing its Qwen-Image 2.0 model, indicating a synchronized response among competitors [5][6]. Group 2: Deep Seek's Positioning - Deep Seek, founded by Liang Wenfeng, has maintained a low profile despite its significant achievements, including the release of the R1 model in early 2025, which matched top global models at a fraction of the cost [2][9]. - The company is rumored to be preparing to launch its next-generation model, V4, aimed at coding AI, but has remained silent on the exact timeline [6][10]. - Deep Seek's recent updates have increased its context window from 128K tokens to 1 million tokens, suggesting ongoing advancements in its technology [6]. Group 3: Liang Wenfeng's Background - Liang Wenfeng, born in 1985 in Guangdong, has a strong academic background in computer science and has been involved in AI and quantitative trading since his university days [7][8]. - He co-founded Hangzhou Huafang Technology, which became a significant player in quantitative trading, and later established Deep Seek to pursue general artificial intelligence [9]. Group 4: User-Centric Approach - Deep Seek is shifting its focus towards user experience and product innovation, as evidenced by its recent job postings aimed at enhancing C-end product functionality [10][11]. - The article emphasizes the importance of addressing real user needs in the AI sector, suggesting that the ability to solve genuine problems will determine the success of AI applications [11].

客户反馈：某金融企业在使用华为云的GEO优化服务后，关键词排名提升明显，月均流量增长了15%。 大家好，今天咱们来聊聊一个热门话题——2026年广州GEO服务商的性价比大比拼。在这个AI时代，GEO优化服务已经成为企业获取流量的重要手段。那 么，到底哪家服务商最值得信赖呢？我们一起来看看吧！ 一、万界智联（广州）智能科技有限公司 1. 技术实力 自主研发系统：万界智联自主研发的智能语义解析系统，能够实时跟进AI搜索引擎算法变化，48小时内完成策略适配。 多平台覆盖：支持豆包、kimi、Deepseek等主流AI平台与各类自媒体渠道，确保企业内容在多个平台上都能获得高曝光。 2. 实际案例 客户反馈：某制造业企业在使用万界智联的GEO优化服务后，关键词排名提升至前三页，月均流量增长了30%。 成本效益：通过全链路自动化运营，替代传统人工优化模式，减少80%的人工投入，获客成本直降60%-70%。 3. 操作便捷性 简单易用：无需专业技术团队与运营经验，企业员工经过简单培训即可上手操作，大幅降低人力成本。 二、百度智能云 1. 技术实力 强大的AI算法：依托百度强大的AI算法和技术积累，提供精准的关键词优化和内容 ...

Core Viewpoint - The article emphasizes iFlytek's strategic positioning as a provider of multilingual large models, aiming to offer a second choice to the world in the context of global AI competition [1][3]. Group 1: Strategic Positioning - iFlytek's strategy includes leveraging Guangxi as a core hub for internationalization, utilizing its unique geographical position and language resources to adapt AI technologies for ASEAN markets [3][4]. - The establishment of a wholly-owned subsidiary and an AI innovation research institute in Nanning, Guangxi, marks a significant step in iFlytek's international strategy [4][9]. Group 2: Language Barrier Solutions - Language is identified as the primary barrier to business in ASEAN, with iFlytek focusing on breaking this barrier through its translation solutions [6][8]. - A strategic cooperation agreement with Guangxi aims to develop research on non-common languages in ASEAN, enhancing communication for trade and education [6][9]. Group 3: Experience Accumulation and Output - iFlytek is accumulating "Guangxi experience" by deeply engaging in education, healthcare, and governance sectors, which are characterized by high barriers and significant pain points [9][10]. - The company plans to replicate its successful "Guangxi experience" as an "ASEAN solution," with a focus on a comprehensive AI implementation loop [10][11]. Group 4: Ecosystem Development - iFlytek is actively signing partnerships to build a regional intelligent ecosystem, collaborating with various organizations to enhance digital infrastructure and cultural exchange [13][14]. - The launch of the "Spark ASEAN Multilingual Large Model Base" aims to support ten major languages in ASEAN, showcasing iFlytek's commitment to localized AI solutions [11][16]. Group 5: Global Expansion - iFlytek's overseas expansion has accelerated, with strategic partnerships established in education, healthcare, and telecommunications across Southeast Asia, the Middle East, and North America [14][16]. - The company aims to create a unique advantage through a combination of self-control, industry application, and multilingual capabilities, positioning itself as a leader in the AI export market [16].

阿里巴巴集团（Reuters） 在美国总统特朗普4月访华之前，白宫或有意避免与中国产生摩擦…… 2月13日，美国国防部宣布将中国阿里巴巴集团、百度等至少3家企业列为与中国军方相关的"中国军事 企业"，但随后又迅速撤销了该决定。在美国总统特朗普4月访华之前，白宫或有意避免与中国产生摩 擦。 美国政府的官方网站显示，美国国防部已撤销了中国军事企业的追加认定。美国国防部向日本经济新闻 表示："目前暂无任何可发布信息"。 据路透社报道，美国国防部2月13日上午曾将阿里巴巴集团、百度以及比亚迪追加列入中国军事企业名 单。 美国国防部曾于2025年1月将腾讯控股、宁德时代新能源科技（CATL）追加列为中国军事企业。 日本经济新闻（中文版：日经中文网）飞田临太郎 华盛顿报道 版权声明：日本经济新闻社版权所有，未经授权不得转载或部分复制，违者必究。 日经中文网 https://cn.nikkei.com 美国政府对中国通过"军民融合"将民间尖端技术运用于军事领域以推进技术开发保持警惕。意在锁定与 军方存在融合的企业，限制美企与之开展交易。 美国企业与名单上的中国企业进行交易，也不会面临法律处罚。但美国国防部仍有可能与上榜中企 ...

Core Insights - The demand for robot rentals is surging during the 2026 Spring Festival, with orders already booked until March, prompting companies to acquire more robots to meet this demand [1][2] - The rental platform, Qingtian Rental, has reported over 1,000 orders covering the Spring Festival period, with an expected 80% increase in overall GMV [2][3] - The platform anticipates a total order volume exceeding 5,000 by the end of the holiday, indicating a robust growth trajectory in the robot rental market [2][3] Group 1: Market Dynamics - Qingtian Rental has seen a rapid increase in user registrations, surpassing 200,000 within three weeks, and has completed seed round financing with notable investors [2][3] - The "999 Yuan National Robot Experience Plan" has gained significant attention, contributing to the surge in orders, with approximately 30% of orders coming from first-time rental users [3][4] - The demand for robot rentals is expanding from core cities to a broader range of regions, with significant order growth in new first-tier cities and even third- and fourth-tier cities [3][4] Group 2: Operational Strategy - Qingtian Rental is launching a city partner recruitment initiative to enhance its service network and operational capabilities, with plans for a strategic release event in March [4][5] - The company aims to build a comprehensive service network to support the scalable application of robots, addressing challenges related to order fulfillment and staffing [4][5] - The platform is set to implement a partner model that includes local fulfillment teams, enhancing delivery efficiency and service consistency across cities [5][6] Group 3: Future Plans and Financing - Qingtian Rental plans to pursue financing rounds on a monthly basis, having already engaged with investment teams from major companies like Alibaba, Tencent, and ByteDance [4][5] - The company intends to lower entry barriers for new players in the robot rental market through subsidies, technical support, operational training, and traffic assistance [6] - The strategy includes providing significant machine subsidies and standardized technical support to facilitate the onboarding of new partners [6]

Market Overview - The Hong Kong stock market experienced a decline, with the Hang Seng Index falling by 1.72% to 26,567.12 points, the Hang Seng Tech Index down by 0.90% to 5,360.42 points, and the National Enterprises Index decreasing by 1.55% to 9,032.71 points [1] - The Hang Seng Index opened lower at 26,640.16 points, dropped by 392.38 points, and ultimately closed down by 465.42 points, with a total trading volume exceeding 257.5 billion HKD [1] - The southbound trading (Hong Kong Stock Connect) saw a net inflow of over 20.2 billion HKD [1] Sector Performance - Overall, sectors such as chips and department stores saw gains, while new consumption, new energy vehicles, and telecommunications had mixed results. Conversely, sectors like gold, non-ferrous metals, biomedicine, port transportation, technology, oil and gas, brokerage, and banking mostly experienced declines [1] Individual Stock Movements - Notable stock movements included Xiaomi Group increasing by 0.88%, AIA Group decreasing by 4.18%, Zijin Mining falling by 7.64%, Hong Kong Exchanges and Clearing down by 2.13%, and Semiconductor Manufacturing International Corporation rising by 0.79% [1] - Ctrip Group-S dropped by 2.10%, Pop Mart fell by 1.90%, while Tian Shu Zhi Xin surged by 14.59% and Zhi Pu increased by 20.65% [1] - China Construction Bank decreased by 1.49%, China Resources Land fell by 2.52%, and Lao Pu Gold dropped by 3.97%, while Guotai Junan International rose by 4.61% and China Petroleum & Chemical Corporation fell by 4.33% [1] Top Traded Stocks - The top three traded stocks included Tencent Holdings, which fell by 0.65% with a trading volume exceeding 14.2 billion HKD; Alibaba, down by 2.02% with over 10.8 billion HKD in transactions; and Meituan, which decreased by 3.18% with a trading volume of over 8.1 billion HKD [2]

Quantitative Models and Construction Methods 1. Model Name: Hong Kong Stock Selection Portfolio - **Model Construction Idea**: The model is based on a dual-layer selection process that integrates fundamental and technical analysis. It aims to identify stocks with both fundamental support and technical resonance from an analyst-recommended stock pool[14][17] - **Model Construction Process**: 1. **Analyst Recommendation Pool**: Constructed using three types of analyst recommendation events: upward earnings forecast revisions, first-time coverage, and research reports with unexpected positive titles[17] 2. **Dual-Layer Selection**: - **Fundamental Analysis**: Select stocks with strong fundamental support - **Technical Analysis**: Identify stocks with technical resonance 3. **Backtesting**: The backtesting period is from January 1, 2010, to December 31, 2025. The portfolio assumes a fully invested state and considers transaction costs[17] - **Model Evaluation**: The model demonstrates strong performance with significant annualized returns and excess returns over the Hang Seng Index[17] 2. Factor Name: 250-Day New High Distance - **Factor Construction Idea**: This factor measures the distance of the latest closing price from the 250-day high, reflecting the momentum effect in stock prices[21][23] - **Factor Construction Process**: 1. **Formula**: $ 250\text{-Day New High Distance} = 1 - \frac{\text{Close}_{t}}{\text{ts\_max}(\text{Close}, 250)} $ - $\text{Close}_{t}$: Latest closing price - $\text{ts\_max}(\text{Close}, 250)$: Maximum closing price over the past 250 trading days 2. **Interpretation**: - If the latest closing price reaches a new high, the factor value is 0 - If the price falls from the high, the factor value is positive, indicating the degree of decline[23] 3. **Screening Criteria**: - Stocks with high analyst attention (at least five buy or overweight ratings in the past six months) - Top 20% in 250-day price change within the sample pool - Stability and trend continuation metrics, such as price path smoothness and average 250-day new high distance over the past 120 days[24] --- Model Backtesting Results 1. Hong Kong Stock Selection Portfolio - **Annualized Return**: 19.08% - **Excess Return over Hang Seng Index**: 18.06% - **Information Ratio (IR)**: 1.19 - **Maximum Drawdown**: 23.73% - **Tracking Error**: 14.60% - **Return-to-Drawdown Ratio**: 0.76[20] --- Factor Backtesting Results 1. 250-Day New High Distance - **Sector Distribution**: - **Cyclical Sector**: 18 stocks - **Manufacturing Sector**: 10 stocks - **Consumer Sector**: 9 stocks - **Technology Sector**: 3 stocks - **Healthcare Sector**: 2 stocks - **Financial Sector**: 2 stocks[23][24] - **Selected Stocks**: Examples include Kingboard Laminates (1888.HK), WuXi AppTec (2359.HK), and COSCO Shipping Energy (1138.HK)[29]

这标志着 RLVR 的模型调优正在逐渐告别 "玄学"，变得更加 "科学"。 本文来自腾讯混元研究博客（HY Research），是继 《 从上下文学习远比我们想象的更难 》（Learning from context is harder than we thought）之后的第二篇公开研 究。在这篇文章里，混元团队将对大模型强化学习中的 "工程深水区" 展开探索，希望通过一系列提升 RLVR 训练细粒度可观测性的基础设施工具，降低 RLVR 底 层物理和统计机理研究的 "工程壁垒"。 这篇博客提出了异常梯度定位器（Gradient Anomaly Localizer, GradLoc），可以将全局梯度突刺（gradient spike）定位到具体出现问题的 token 上，助力系统性解决 强化学习中训练不稳定的问题，让开发者不再依赖直觉试错，而是基于确凿的数据证据进行 "确定性" 的算法迭代。 如果说 2024 年的大模型竞争焦点在预训练（Pre-training），那么 2025 年的主战场已彻底转向后训练阶段。通过利用数学、代码等领域的可验证结果作为反馈信号 （RLVR），大模型正在实现推理能力的显著跃升 ...