Group 1 - SenseTime launched Seko 2.0, the industry's first multi-episode generative AI model, which is based on its self-developed Seko series model. The model has successfully adapted to the domestic AI chip Cambricon, with a strategic cooperation established in October to optimize software and hardware integration [2] - Former OpenAI CTO Mira Murat has founded Thinking Machines Lab, with the latest product Tinker now valued at $50 billion. The product features the trillion-parameter Kimi K2 Thinking model, designed for long-duration reasoning and tool invocation [2] - Mill Electronics introduced the RK3576 AI edge computing box, which offers high computing power, low power consumption, and strong scalability, becoming a key tool for upgrading industrial vision, engineering machinery, and smart city sectors [2] Group 2 - Google integrated NotebookLM into Gemini, enhancing user interaction by allowing users to attach notes for additional context during conversations with AI chatbots [2]

| 机器之心编辑部 | | --- | | 当前，AI 领域的研究者与开发者在关注 OpenAI、Google 等领先机构最新进展的同时，也将目光投向了由前 OpenAI CTO Mira Murati 创办的 Thinking Machines | | Lab。 | | 今年早些时候，他们推出了首款产品 Tinker ：这是一个 API，用于帮开发者 / 研究人员微调语言模型。你只需要专注于训练数据和算法，而你不擅长的关于 Infra | | 的部分 —— 调度、调优、资源管理和 Infra 可靠性 —— 统统由 Tinker 来搞定，从而大大简化了 LLM 的后训练过程。 | | 此前，Tinker 仅向研究人员和开发者开放内部测试；而如今，Thinking Machines 宣布 正式取消候选名单，所有用户都可以直接使用 Tinker 。 | | 除此以外，Tinker 还带来了其他三项更新： | | 首先，更强推理能力：用户现在可以在 Tinker 上 对 Kimi K2 Thinking 进行微调 。 Kimi K2 拥有万亿参数规模，是 Thinking Machines 目前规模最大的模型，专为 ...

China (PRC) | Technology Equity Research China AI in a Global Context — A Global "Power" Struggle 1) GOOG's Gemini 3 overtook GPT5 to be No 1 in model performance. 2) Moonshot's Kimi K2 overtook MiniMax M2 as the best-performing LLM in China, only 8% < Gemini 3. 3) ZTE launched a smartphone powered by ByteDance's Doubao Mobile Assistant, a good demo of how AI could enhance a smartphone but commercial success unlikely. 4) The US allows NV H200 chips to be sold to China but if China wants it is uncertain. We ...

Core Viewpoint - The article discusses the rise of Zhang Yutong, a prominent figure in the AI startup "Moon's Dark Side," highlighting her transition from investor to CEO and the company's rapid valuation growth as it prepares for a potential IPO by 2026 [1][10]. Group 1: Zhang Yutong's Background and Role - Zhang Yutong, a Tsinghua University graduate and former partner at Sequoia Capital, has a notable investment history, including investments in high-profile projects like Xiaohongshu, which is valued over $31 billion [3][4]. - Her appointment as CEO marks a significant shift from being a behind-the-scenes investor to leading the company's operations and strategy [4][5]. Group 2: Controversies and Disputes - A key controversy involves Zhang's departure from Sequoia Capital after securing over $1 billion from Alibaba for Moon's Dark Side, which increased the company's valuation from $300 million to $2.5 billion [5][6]. - Former colleague Zhu Xiaohu has publicly criticized Zhang, alleging she concealed important information regarding her equity stake in the new venture, which has led to ongoing disputes [8][9]. Group 3: Capital Market Activity - Moon's Dark Side has rapidly raised over 3 billion yuan in five funding rounds since June 2023, attracting major investors like Sequoia China and Tencent, with its valuation soaring from nearly 2 billion yuan to approximately $2.5 billion [11][12]. - The latest funding round is expected to push the company's valuation to around $4 billion (approximately 28 billion yuan) [12]. Group 4: Market Position and Challenges - Despite significant capital influx, Moon's Dark Side faces challenges in user engagement, ranking sixth among AI assistants with about 9 million active users, trailing behind competitors [13]. - The company aims to launch a new generation of its AI model, Kimi K2 Thinking, which promises to enhance its technological capabilities and address commercialization challenges [13].

作者 | 周智宇 编辑 | 张晓玲 张予彤，这位一度引起争议的金沙江创投前主管合伙人，以一个全新身份走向台前。 近日真格基金在清华大学举办的一场交流会上，张予彤首次以"Kimi总裁"的身份公开亮相。她负责的是 Kimi整体战略与商业化。 张予彤也借着这场演讲，回应了外界对于"独角兽资金不足、算力匮乏"的质疑，强调Kimi的效率优势。 从某种程度上来说，这也是场另类路演。 放眼望去，曾经并肩作战的"大模型六小虎"已在分岔路口渐行渐远：抢滩上市的急迫、无奈折叠万亿参 数雄心的妥协，以及被价格屠夫无情击穿底线的恐慌，共同交织成一幅残酷的众生相。在巨头围剿与资 本退出的双重夹击下，所有的技术信仰最终都必须兑换成财务报表上的数字。 张予彤走向台前，正是月之暗面试图穿越这片商业"无人区"的最后一搏，也预示着这场关乎生死的中场 战事，来到重要赛点。 走向台前 杨植麟需要张予彤。或者更准确地说，处于"中场战事"的月之暗面，急需一位懂资本、懂战略、更懂如 何把技术兑换成商业价值的操盘手。 这是一场跨越十年的重逢，也是一次角色的彻底重塑。作为清华系的"师姐"，张予彤曾是杨植麟上一家 创业公司循环智能的伯乐。 如今，她正式成为这家 ...

新征程。 作者/ 周佳丽 吴琼 报道/投资界PEdaily 近日，张予彤意外出现在清华大学的一场交流会上。 不同于以往，这一次她带着新t i t l e亮相——月之暗面Kimi总裁，"算是第一次以这个职 务身份亮相"。 投资界从接近 Ki m i 人士了解到，张予彤已经出任月之暗面总裁一职，"负责公司的整 体战略与商业化，包括融资，也会参与一些新产品的开发。" 就这样，杨植麟邀请张予彤作为联合创始人加入月之暗面。 根据 他此前的声明，张予 彤的股份按照多年兑现（v e st i n g），兑现的条件是持续性为公司提供多年的服务及产 出业绩。 此 后 ， 创 投 圈 见 证 了 月 之 暗 面 的 融 资 速 度 ， 身 后 集 结 了 红 杉 中 国 、 真 格 基 金 、 砺 思 资 本、今日资本等知名基金以及阿里、美团、小红书等大厂，估值也是螺旋式上升，早已 挺进3 0亿美元大关。 这当中，张予彤起到了不可或缺的作用。尤其是月之暗面阿里融资案中，她被认为是背 后最重要的推动者。杨植麟此前也在一份声明中提到，（张予彤）在业务、战略以及多 场融资战役中对公司做出了重要贡献。 回想过去一年，张予彤与老东家金沙江 ...

Core Insights - The latest xbench-ScienceQA leaderboard has been released, showcasing new models from six companies, with Gemini 3 Pro achieving state-of-the-art (SOTA) performance and DeepSeek V3.2 matching GPT-5.1 in scores while offering high cost-effectiveness [1][2][6] - xbench will introduce two new benchmarks to evaluate agent instruction-following capabilities and multimodal understanding of models [1] Model Performance Summary - **Gemini 3 Pro**: Scored 71.6, up from 59.4 in Gemini 2.5 Pro, with a BoN of 85. Average response time is 48.62 seconds. Cost for answering 500 questions is approximately $3 [3][6] - **DeepSeek V3.2**: Achieved a score of 62.6, matching GPT-5.1, with a BoN of 81. The cost for 500 questions is only $2 for the Speciale version and $1.3 for the Thinking version [6] - **Claude Opus 4.5**: Scored 55.2 with a fast average response time of 13 seconds, showing improvement over its predecessor [6] - **Kimi K2 Thinking**: Scored 51.8 with a BoN of 76, indicating a slight improvement [6] New Model Developments - **DeepSeek V3.2**: Introduces a Sparse Attention mechanism to enhance long-context performance while reducing computational complexity. It also features a scalable reinforcement learning framework to improve reasoning and instruction-following capabilities [10][12] - **Gemini 3**: A new multimodal model from Google DeepMind, excelling in reasoning depth and multimodal understanding, achieving a top score of 1501 Elo in LMArena [13] - **Nano Banana Pro**: A new image generation model that integrates advanced reasoning capabilities with real-time knowledge, allowing for complex image synthesis [14] - **Claude Opus 4.5**: A flagship model from Anthropic that excels in code generation and human-computer interaction, achieving high performance in real-world software engineering tasks [15][16] - **GPT-5.1**: An important iteration from OpenAI that enhances conversational fluency and complex task reasoning, introducing adaptive reasoning mechanisms [17] - **Tongyi DeepResearch**: Designed for deep research tasks, this model combines mid-training and post-training frameworks to enhance agent capabilities, achieving competitive performance with a smaller model [19]

Core Insights - The AI unicorn "Moon's Dark Side" is nearing the completion of its latest funding round, with a valuation expected to rise to approximately $4 billion [1] - The company aims to initiate an IPO in the second half of next year after securing several hundred million dollars in funding [1][2] Company Overview - "Moon's Dark Side" was established in April 2023, founded by Yang Zhilin, a Tsinghua University graduate with a PhD from Carnegie Mellon University [2] - As of January this year, the company's valuation reached $3.3 billion [2] Product Development - The latest model, Kimi K2 Thinking, has surpassed well-known models like GPT-5 and Claude 4.5 in key benchmark tests [2][3] - K2 Thinking achieved state-of-the-art (SOTA) performance in several tests, including "Humanity's Last Exam" and complex information gathering [3] Future Plans - The team is planning significant architectural changes for the upcoming K3 model, focusing on enhancing performance in long-sequence tasks [4] - The company remains committed to developing foundational large models, unlike many startups that have shifted away from this path [6] Industry Context - Competitors like Zhiyuan and MiniMax are also pursuing IPOs, with Zhiyuan being the first to initiate the IPO process among large model startups [5][6] - The competitive landscape is intensifying, with major players like ByteDance, Alibaba, and Tencent investing heavily in the AI sector [6]

Core Insights - The AI unicorn "Dark Side of the Moon" is nearing the completion of its latest funding round, with a valuation expected to rise to approximately $4 billion [1] - The company aims to initiate an IPO in the second half of next year after securing several hundred million dollars in funding [1] - The competitive landscape for large models is intensifying, with "Dark Side of the Moon" regaining market attention following the release of its latest model, K2 [1][2] Company Developments - "Dark Side of the Moon" was founded in April 2023 by Yang Zhilin, a Tsinghua University graduate with a PhD from Carnegie Mellon University [1] - The company’s valuation reached $3.3 billion as of January this year [1] - The latest model, Kimi K2 Thinking, has surpassed notable models like GPT-5 and Claude 4.5 in key benchmark tests [2] Model Performance - Kimi K2 Thinking achieved state-of-the-art (SOTA) performance in several benchmarks, including "Humanity's Last Exam" and complex information retrieval tasks [2] - The model is capable of executing 200 to 300 tool calls to solve complex problems, ensuring task continuity [2] Future Plans - The team is planning significant architectural changes for the upcoming K3 model, potentially adopting a new design philosophy based on their experiments with the KDA architecture [3] - Other companies in the industry, such as Zhiyuan and MiniMax, are also pursuing IPOs, indicating a trend among large model startups [3] Industry Context - The large model sector is characterized by fierce competition, with major players like ByteDance, Alibaba, and Tencent investing heavily to secure their positions [5] - "Dark Side of the Moon" remains one of the few companies continuing to invest in foundational large models despite a trend of startups moving away from this path [4]

Group 1: Google Product Launches - Google has launched two significant products: the Gemini 3 model and the AI-native IDE product Antigravity, marking a new era in AI and a step towards AGI [4][5] - Gemini 3 has achieved remarkable performance, surpassing competitors in various benchmark tests, and is integrated directly into Google Search [5][7] - The model demonstrates superior reasoning and multimodal understanding capabilities, making it effective for complex decision-making tasks, and is offered at a lower price point compared to competitors [5][8] Group 2: AI Developments and Trends - OpenAI has released the GPT-5.1 version, enhancing intelligence and user interaction by allowing customization of AI personality and tone [9][10] - The new version includes six preset dialogue modes and adaptive reasoning capabilities, aiming to create a more personalized user experience [11][13] - Manus has introduced the Browser Operator extension, enabling any browser to function as an AI assistant without requiring new applications or configurations, thus lowering the barrier for AI integration [12][14][15] Group 3: Industry Insights - The McKinsey 2025 AI Report reveals that while most companies are adopting AI, only a third have achieved scalable applications, highlighting a gap in effective utilization [16][17] - The report emphasizes that AI should not merely be a tool but a catalyst for innovation, urging companies to integrate it deeply into their core business processes [17] - Kimi K2 Thinking has emerged as a leading open-source model, outperforming previous benchmarks and redefining standards in the open-source AI sector [18][19] Group 4: Performance Upgrades - Grok 4 Fast has undergone a significant upgrade, expanding its context window to 2 million tokens, enhancing real-time AI reasoning capabilities [20][21] - The accuracy of reasoning modes has improved significantly, indicating a leap in performance quality for AI models [21]