21世纪经济报道记者董静怡 2026年春节前夕，MiniMax正式上线新一代文本模型MiniMax M2.5，于2月12日在 MiniMax Agent上 线，并于13日全球开源支持本地化部署。 2月12日，MiniMax涨幅一度超过20%。12日收盘，MiniMax股价报588港元/股，涨幅达14.62%。13日， MiniMax开盘继续大涨，截至午盘，涨10%，总市值超2000亿港元。 据介绍，在编程能力方面，M2.5在权威榜单SWE-Bench Verified得分80.2%、Multi-SWE-Bench得分 51.3%，较上一代显著提升；在Multi-SWE-Bench等多语言复杂环境中超越Opus 4.6，达到了行业最好的 水平。 更重要的是，模型展现出"原生Spec能力"——在编码前主动拆解架构与功能规划，更接近真实架构师的 工作模式。 M2.5-lightning版本支持100 TPS以上输出速度，是主流模型的2倍左右；输入价格约0.3美元/百万 Token，输出约2.4美元/百万Token。按每秒输出100 Token计算，连续运行一小时成本约1美元；若按50 Token计算，成本约0.3美 ...

中国AI行业的"百模大战"正在落幕，真正的玩家已缩减至10家。最刺耳的结论是，最大利润池将归掌控分发的腾讯、阿里 等大厂，而非模型公司。独立厂商中，智谱靠本地化部署守住59%高毛利，MiniMax则靠73%的海外收入和全模态产品突 围。当模型不再稀缺，变现才是王道。 硬·AI 作者 | 龙 玥 编辑 | 硬 AI 当模型不再稀缺，真正稀缺的，是能把模型变成现金流的能力。 2月9日，摩根大通证券（中国）发布研究报告《中国人工智能行业：全球布局与模型创新驱动新一代领军 者》，首次覆盖中国独立大模型厂商智谱与MiniMax。 报告开宗明义：中国人工智能行业正从"百模大战"阶段，迈向以商业化落地能力、模型创新实力及全球化 布局为决定成败之关键的阶段。中国AI市场正在迅速整合， "具备实力且资金充足的模型开发商数量已从 超200家缩减至不足10家。" 摩根大通犀利指出， 国内AI行业最大的利润池恐将流向掌握分发的平台巨头 ；而独立厂商的突围，则取 决于谁能通过"结构性中立"找到生存缝隙——智谱向内深耕高合规的本地化部署，MiniMax向外拓展高溢 价的全球市场。 这一判断的背景并不复杂。报告指出，随着模型训练成本、算力 ...

中国AI行业的"百模大战"正在落幕，真正的玩家已缩减至10家。最刺耳的结论是，最大利润池将归掌 控分发的腾讯、阿里等大厂，而非模型公司。独立厂商中，智谱靠本地化部署守住59%高毛利， MiniMax则靠73%的海外收入和全模态产品突围。当模型不再稀缺，变现才是王道。 当模型不再稀缺，真正稀缺的，是能把模型变成现金流的能力。 2月9日，摩根大通证券（中国）发布研究报告《中国人工智能行业：全球布局与模型创新驱动新一代 领军者》，首次覆盖中国独立大模型厂商智谱与MiniMax。 在摩根大通看来，这一阶段的核心变化在于：模型能力逐步趋同；资金消耗呈指数级上升；客户开始 更关注"交付能力、稳定性与可持续性"。 这意味着， 大模型竞争的主线，正在从技术竞赛，转向商业系统的构建能力。 | 股票评级和目标价 | | | | | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | 市值 | | 评级 | | | 目标价 | | | 公司名称 | 代码 | (百万美元) 货币 | 股价 | 新 | 日 | 新 | 截至日期 | le | | ...

Group 1: Core Insights - Talent density is a key factor in capital pricing within the AI sector, with high valuations reflecting the future value that talent can create [2] - The second wave of talent migration in AI was triggered by the resurgence of interest in the industry following the launch of ChatGPT in 2022 [4][5] - The "AI Six Dragons" have attracted over 10 billion in capital investments between 2023 and 2024, with companies like Zhipu and MiniMax seeing valuations exceed 20 billion [5][11] Group 2: Talent Movement - High-end talent is increasingly flowing back to major tech companies, with significant salary increases reported, such as algorithm engineers receiving up to a 30% salary boost or even doubling their pay when moving to larger firms [11] - The talent flow from the "AI Six Dragons" indicates a trend where technical roles are returning to big companies, while product and business roles are more inclined to pursue entrepreneurial ventures [6][11] - Major companies like ByteDance, Tencent, and Alibaba are actively recruiting top AI talent, with reports of salaries reaching up to 10 million for new graduates [10][11] Group 3: Competitive Landscape - The "Hundred Model War" began in March 2023, with various companies rapidly iterating and releasing new models, reflecting the competitive nature of the AI landscape [12][16] - By the end of 2023, the "AI Six Dragons" had expanded significantly, with Zhipu AI growing to over 400 employees, 70% of whom are in R&D [20] - ByteDance's strategic entry into the AI market, including the launch of its Doubao model, marked a significant shift in the competitive dynamics among AI companies [21][24]

Core Viewpoint - Baidu's Wenxin Assistant has surpassed 200 million monthly active users, positioning itself alongside ByteDance's Doubao and Alibaba's Qianwen as one of the "AI three giants" [3][15]. Group 1: User Engagement and Strategy - Baidu processes over 600 million search requests daily, integrating AI into its search platform to enhance user experience without requiring users to download new applications [4][16]. - The strategy of embedding AI into the search experience allows Baidu to avoid direct competition with Doubao and Qianwen, focusing instead on the "first scene of demand" where users naturally engage with AI [5][17]. - The success of reaching 200 million monthly active users is attributed to the natural accumulation of users from the existing search traffic rather than aggressive subsidies [5][17]. Group 2: Competitive Landscape - The competitive landscape has shifted from a focus on technical specifications to who can effectively serve high-frequency scenarios [6][18]. - Baidu's advantage lies in its ability to meet users' explicit information needs, as opposed to casual interactions seen in other platforms [7][18]. - The integration of various tools within the search ecosystem allows for a seamless user experience, enhancing retention through practical applications [7][18]. Group 3: Challenges and Concerns - Despite the impressive user numbers, there are concerns regarding the quality of engagement, as many interactions are single-use and lack user stickiness compared to Doubao's average of five interactions per user [8][19]. - Baidu's ecosystem lacks strong transactional capabilities compared to Alibaba and ByteDance, making it difficult to convert user interactions into revenue [8][19]. - The perception of Baidu's technology remains weak among users, with lingering doubts about the effectiveness of Wenxin Assistant despite its technical achievements [8][19]. Group 4: Future Outlook - The ongoing competition is fundamentally about "scene sovereignty," with each player dominating different user engagement scenarios [10][21]. - Future developments could threaten Baidu's position if competitors like WeChat and Douyin expand their AI capabilities [11][22]. - The 200 million monthly active users mark a starting point for Baidu, with the real challenge being to transition users from passive engagement to active utilization of AI services [12][23].

责任编辑：李思阳 专题：未竟之约：张小珺访谈录 近日在《未竟之约》栏目中，智谱AI CEO张鹏在与张小珺对话中谈到"2023年百模大战"时表示，当时 有两种感觉：第一感觉是特别兴奋。迎来了一个很大的机会，很大的浪潮，大家都不用再去教育投资 人、教育市场。"反正你做这个事情，大家一听，很炸，OpenAI干这个事吧？行，挺好的。"他说。 但他也直言，此外还是有些焦虑、有些担心。因为每逢大浪过来，往后可能是一片狼藉，最后留不下点 啥，害怕这种状况。 新浪声明：所有会议实录均为现场速记整理，未经演讲者审阅，新浪网登载此文出于传递更多信息之目 的，并不意味着赞同其观点或证实其描述。 但他也直言，此外还是有些焦虑、有些担心。因为每逢大浪过来，往后可能是一片狼藉，最后留不下点 啥，害怕这种状况。 新浪声明：所有会议实录均为现场速记整理，未经演讲者审阅，新浪网登载此文出于传递更多信息之目 的，并不意味着赞同其观点或证实其描述。 责任编辑：李思阳 专题：未竟之约：张小珺访谈录 近日在《未竟之约》栏目中，智谱AI CEO张鹏在与张小珺对话中谈到"2023年百模大战"时表示，当时 有两种感觉：第一感觉是特别兴奋。迎来了一个很大的机 ...

Core Viewpoint - The AI application market is undergoing a silent cleansing, with a significant number of companies exiting the market due to high operational costs and lack of sustainable business models [4][5][18]. Group 1: Market Dynamics - Since the launch of ChatGPT until July 2024, 78,612 newly registered AI companies in China have disappeared, accounting for 8.9% of the total new registrations during the same period [5]. - The AI application industry is experiencing a bifurcation, where only products that understand long-term operations can survive through cycles [7]. - The rapid expansion of applications like Quark and Doubao contrasts sharply with the mass exit of companies, indicating a market correction as capital enthusiasm wanes [6][7]. Group 2: AI Entrepreneurship Trends - The AI entrepreneurship wave is driven by a mix of genuine demand and market anxiety, with many companies rushing to enter the AI space without a clear understanding of user needs [9][13]. - The cost of using large models has significantly decreased, with models like GPT-3.5-turbo costing only $0.002 per 1,000 tokens in 2023, a 90% reduction compared to previous versions [10]. - The proliferation of AI applications has led to a situation where almost anyone can leverage large models for entrepreneurship, but many lack the necessary technical maturity [11][13]. Group 3: Challenges Faced by AI Products - Many AI products are failing due to unclear business models and a lack of paying user bases, leading to a cycle of high technical investment and weak revenue generation [22][23]. - The high degree of product homogeneity in the AI space has resulted in a lack of core competitiveness, with many products relying on third-party APIs rather than developing unique technologies [25][27]. - The intense competition in the AI market, exacerbated by the entry of major tech giants, has squeezed the survival space for smaller players, leading to price wars and unsustainable growth strategies [32][33]. Group 4: Strategies for Long-term Viability - Successful AI applications need to cover a broad range of market demands and lower usage barriers, ensuring a wide user base [37]. - Rich product functionality is essential, as single-function applications often struggle to attract a large user market [37][38]. - Advanced technology and diverse data sources are critical for building core competitiveness, allowing applications to create unique user value and maintain relevance in a crowded market [40][41]. Conclusion - The exit of 80,000 AI application companies does not negate the value of AI but highlights the need for products that solve real problems and create tangible value [43]. - The future of AI belongs to those who focus on practical solutions rather than chasing trends [44].