文 | AIX财经，作者 | 王璐，编辑 | 魏佳 在智谱AI披露招股书之后不到48小时，同属"大模型六小龙"的MiniMax（上海稀宇科技有限公司）也公 布了招股书。就在几天前，两家公司已先后通过港交所上市聆讯。两份招股书前后脚亮相，将国内头部 大模型公司的商业路径、财务数据与潜在风险，首次摆到大众面前。 智谱AI的招股书率先引发关注，其持续高投入的研发模式与三年半累计亏损超62亿的金额，成为市场 讨论焦点。外界普遍将其视为深耕底层技术的代表，商业化主要围绕B端客户展开。 MiniMax则呈现出另一条发展轨迹。过去三年零九个月里，其营收8742万美元（约合6.2亿元人民 币），不及智谱AI三年半的表现，但同期累计亏损达13.2亿美元（约合92.9亿元人民币），超过智谱 AI。与此同时，MiniMax在招股书中着重强调了其多模态大模型能力，以及由此孵化出的多款C端产品 矩阵，收入结构更偏向AI产品侧变现。 两家公司的路线差异，让外界的讨论进一步分化。一部分人认为，在目前大模型公司都处在亏损的现状 下，MiniMax"多模态+重产品"的打法，对资金与算力的消耗更为集中，C端应用能否持续覆盖成本仍存 在不确定性。 ...

Tech Ideas 线 | 讨论会 12 曲凯 42章经创始人 莫傑麟 家族办公室资深从业者 核心议题 25 年－二级市场 年终复盘和 26 年展望 关键词 多月不 AT 硬件 Agent 他熟悉硅谷二级市场，我深耕国内一级市场。这种跨视角的对照，使得我们每次凑在一起，往往能碰撞出一些相对超前、也还蛮准确的结论（不少听众反馈，我 们在节目里的很多判断，后来都应验了）。 过去这一年，我们录了三期播客： 《世界怎么就「东升西落」了？聊聊二级市场与 DeepSeek+Manus 的热潮》 《硅谷 AI 大转弯与二级市场的牛市》 《「你觉得 AI 有泡沫吗？」——有》 从今年秋天开始，我们又把这种季度复盘，延展成了一个更小范围、更高密度的形式：Tech Ideas 线上讨论会。 每一期，都会由我、莫傑麟，以及几位长期研究产业 / 投资的朋友共同主持。我们会围绕近期的重点主题，邀请业内朋友们，一起进行小范围的交流讨论。在上一 期活动中，嘉宾的分享中也有不少有启发性、扎实的 insights。 今年的最后一场活动，我们想做一次更完整的收官与展望：复盘 2025 年的一、二级市场，并展望 2026，共同探讨今明两年的 A ...

Core Insights - The AI market in China is rapidly evolving, with Huoshan Engine emerging as a leading player, particularly in the model-as-a-service (MaaS) sector, where it holds the largest market share domestically and ranks third globally [2][3] - The daily token usage of the Doubao model has surged to over 50 trillion, marking a tenfold increase compared to the previous year [1][4] - The focus for 2025 in the AI market will be on multimodal capabilities and agents, with Huoshan Engine launching several new products centered around these themes [3][6] Market Position and Growth - Huoshan Engine has established itself as a significant force in the AI sector, with projected revenues exceeding 200 billion in 2024, reflecting a growth rate of over 60% [6][23] - The company aims to simplify model usage by integrating multiple capabilities into a single API, contrasting with competitors who offer separate models for different functions [26][27] Product Innovations - The newly launched Seedance 1.5 pro video generation model emphasizes immediate usability, capable of producing synchronized audio-visual content without extensive post-production [8][15] - The model's advancements include improved lip-sync accuracy and enhanced immersion, making it particularly suitable for diverse content creation [13][21] Competitive Landscape - The AI video model market is characterized by rapid iteration, with companies focusing on producing fully publishable works rather than just raw video segments [7][9] - Huoshan Engine's approach to model training and optimization has led to a tenfold increase in inference speed, significantly reducing costs and enhancing performance [31][30] Future Directions - The company is exploring innovative billing models, such as the "AI Savings Plan," which offers tiered discounts to help businesses reduce costs by up to 47% [32][33] - Huoshan Engine is committed to building a comprehensive AI infrastructure that enables businesses to easily adopt advanced AI capabilities, aiming to make AI assistants as ubiquitous as websites and apps [38][39]

Investment Rating - The industry investment rating is "Positive" and maintained [6] Core Insights - The report highlights a significant trend in downstream demand for AI applications, driven by the recent launch of the Doubao model 1.8 and the Seedance 1.5 pro video creation model at the Huoshan Engine's Winter Force Conference [2][4] - The Doubao model's daily token usage has surged to over 50 trillion, marking a 471-fold increase since its launch and more than a tenfold increase year-on-year, indicating strong demand across various industries [9] - The introduction of a "savings plan" for models, offering discounts of up to 47%, aligns pricing strategies with customer usage patterns, enhancing affordability and encouraging innovation [9] Summary by Sections Event Description - On December 18, Huoshan Engine held the Winter Force Conference, where the Doubao model 1.8 and Seedance 1.5 pro were officially launched, sparking extensive market discussions [4] Event Commentary - The report emphasizes the explosive growth in the usage of the Doubao model, reflecting genuine customer needs and the model's ability to empower various sectors [9] - The Doubao model 1.8 features enhanced multimodal capabilities, including increased video frame understanding and improved agent functionalities, which are expected to unlock more application scenarios [9] - The conference also introduced several upgraded AI agent products aimed at delivering tangible value to enterprises, such as the AgentKit platform and various specialized agents [9] - The report anticipates a further increase in industry token usage next year, particularly in multimodal applications and edge devices [9]

2025.12. 18 这场由巨头博弈引发的震荡，将一个核心议题推至台前：在以大模型为核心的AI时代，硬件的技术范 式是否正在从通用GPU转向专用芯片如TPU？这是否意味着一场结构性的变革已然来临？ 本文字数：1632，阅读时长大约3分钟 作者 | 第一财经 刘佳 这一悬念不仅关乎国际巨头的战略布局，也紧密牵动着中国AI算力产业链的神经。作为对标英伟达、 不久前刚刚上市的中国GPU厂商代表，摩尔线程创始成员、摩尔学院院长李丰与沐曦高级副总裁孙 国梁在今日腾讯contech大会上"同框"，并回应了对于两种路线的看法。 在李丰看来，争议背后其实是"通才与专才"的分工，而非简单的替代关系。 他分析，谷歌能做TPU，本质上是因为它是全栈整合公司。谷歌有强大的 Infra、基础模型与云服务 形成闭环，把模型跑在自家芯片上量身优化，实现成本性价比的最大化。"但绝大部分企业不具备这 样的垂直整合能力。" 他总结，GPU持续保持优势的原因有三个：灵活度是"甜点"、多模态时代的全功能性、生态的护城 河。 谷歌新一代AI模型Gemini 3系列的发布，在硬件领域投下一颗"重磅炸弹"——其自研TPU（张量处 理器）所展现的性能与成 ...

机器之心发布 如果把用户在互联网上留下的每一个足迹都看作一段记忆，那么现在的推荐系统大多患有 "短期健忘症"。 受限于算力和存储，那些沉睡在数年前的点击、收藏与购买，往往被粗暴地截断或遗忘。即便被召回，它们在模型眼中也只是一串串冰冷且互不相识的 ID 代码。 但事实上，真正有趣的东西也往往藏在这些被遗忘的 "长尾" 之中。如何唤醒这 10 万级 的沉睡数据，并读懂它们背后的视觉与语义关联？ 阿里妈妈与武汉大学团队给出的答案是 MUSE（MUltimodal SEarch-based framework） 。这不仅仅是一个新的 CTR 模型，更像是一个给推荐系统安装的 "多模 态海马体"。它利用图像与文本的语义力量，重构了用户跨越时空的兴趣图谱。 甚至，他们还开源了构建这个 "数字大脑" 的基石： Taobao-MM 数据集 。 对于推荐系统长久以来技术演进路线，这一突破可谓是一次深刻的反思与重构！ 论文标题：MUSE: A Simple Yet Effective Multimodal Search-Based Framework for Lifelong User Interest Modeling 在搜推 ...

这些技术推动大模型从"AIforX"（为某领域提供工具）向"AIinX"（深度嵌入业务流程）演进，为构建 能与物理环境持续交互、自我演进的智能体奠定基础。 【环球网科技报道 记者 张阳】2025年12月9日，由商汤科技与香港科技园公司联合主办的"2025商汤科 技AI论坛"在香港科学园举行。以"模型智未来"为主题，论坛聚焦大模型演进、多模态融合、具身智能 落地及AI驱动的产业范式转移，吸引全球AI研究者、企业决策者与生态伙伴共议技术拐点与商业路 径。 从感知到具身： AI 进入 " 物理世界交互 " 新阶段 商汤科技CEO徐立在开场致辞中指出："我们正经历史上最大规模的技术浪潮——AI已从感知走向生 成，从云端走向端侧，并加速迈向具身智能与世界模型。"作为一家根植香港、服务全球的AI原生企 业，商汤将持续连接国家AI战略与全球创新网络。 香港科技园CEO黄秉修则强调，商汤的成长印证了香港具备孕育世界级科技企业的土壤，"其从初创到 全球领先的跃迁，正是我们支持硬科技创业的最佳案例。" 多模态大模型进入 " 价值闭环 " 攻坚期 商汤联合创始人、首席科学家林达华教授认为：经过三年爆发式发展，行业站在新十字路口。" ...

Core Viewpoint - SenseTime officially launched Seko 2.0, the industry's first multi-episode generative intelligence agent, leveraging its self-developed Seko series models [1] Group 1: Product Development - The Seko series models have completed adaptation to the domestic AI chip Cambricon, marking a significant advancement in supporting AIGC core scenarios from language to multi-modal capabilities [1] - Following the adaptation, SenseTime and Cambricon will further collaborate on deep optimization across multiple directions [1]

Core Viewpoint - The emergence of DeepSeek has significantly impacted MiniMax and other large model companies, prompting introspection on their performance and strategic choices [5][6]. Group 1: Challenges and Reflections - MiniMax's founder, Yan Junjie, faced numerous challenges during the startup phase, including the bankruptcy of Silicon Valley Bank, which affected payroll [3]. - The team recognized that their performance was hindered by a lack of deep thinking and lowered expectations, contrasting with DeepSeek's unique insights and technical accumulation [6][8]. Group 2: Team Morale and Incentives - To boost team morale during tough times, Yan emphasized the importance of encouragement and financial incentives, stating that monetary rewards are effective [7]. - In September, MiniMax initiated a million-dollar stock option incentive program, offering varying amounts based on employee contributions, covering various roles within the company [7]. Group 3: Strategic Direction - MiniMax's approach involves a unique strategy of ToC (Technology of Communication) and international expansion, with their Talkie application gaining significant user traction overseas [8]. - The company experienced a period of indecision regarding whether to prioritize technology or product development, ultimately deciding on a technology-driven approach despite the associated risks [8][9]. Group 4: Market Position and Talent - The gap between domestic large model companies and top international models is narrowing, with Chinese companies achieving this with significantly lower investment [12]. - Yan highlighted the importance of local AI talent, noting that many key contributors to success in companies like DeepSeek and MiniMax are homegrown, often in their first jobs [12]. Group 5: Future Outlook - Yan remains optimistic about the future of AGI, noting that the number of companies in the large model space is decreasing, leading to a more concentrated market [13]. - The AI industry is not merely an extension of the internet; the core product in the large model era is the model itself, with blurred boundaries between roles in product management, development, and algorithms [14].

Core Viewpoint - The article discusses Meituan's strategic focus on AI infrastructure amidst intense competition in the food delivery market, highlighting the company's aggressive approach to AI development and application [3][5]. Group 1: AI Talent Acquisition and Leadership - Pan Xin, a former partner at Flash Technology and head of visual model AI platform at ByteDance, has joined Meituan to lead multimodal AI innovation [4]. - Meituan's AI strategy is built on three levels: AI at work, AI in products, and building large language models (LLMs) [6]. Group 2: AI Model Development and Applications - Since 2025, Meituan has made significant progress in developing foundational models and applications, completing a multimodal foundation that includes language, visual, audio, and video capabilities [8]. - In October, Meituan launched AI tools like "Kangaroo Advisor" and "Smart Manager" for restaurant merchants, making them available for free to all industry players [9]. Group 3: Recruitment and Business Focus - Meituan has been actively recruiting AI talent, particularly in model training, with high standards for candidates primarily sourced from Alibaba, Tencent, and other leading tech firms [7]. - The company is shifting its focus from independent consumer-facing AI applications to integrating AI into its core business operations [11].