（文章来源：第一财经） 智谱宣布GLM-4.7上线并开源。新版本面向Coding场景强化了编码能力、长程任务规划与工具协同，并 在多项主流公开基准测试中取得开源模型中的领先表现。目前，GLM-4.7已通过BigModel.cn提供API， 并在z.ai全栈开发模式中上线Skills模块，支持多模态任务的统一规划与协作。 ...

Core Viewpoint - The release of GLM-4.7 by Zhiyu marks a significant advancement in open-source AI models, achieving top performance in various benchmark tests and surpassing GPT-5.2 in a global coding evaluation system [1][1]. Performance Highlights - GLM-4.7 achieved the best performance among open-source models in multiple mainstream benchmark tests [1]. - In the Code Arena, a prestigious coding evaluation system with participation from millions of users, GLM-4.7 ranked first among open-source models and first among domestic models, outperforming GPT-5.2 [1][1]. Feature Enhancements - The model has enhanced coding capabilities, long-range task planning, and tool collaboration, specifically targeting coding scenarios [1]. - Improvements have also been made in areas such as chat, writing, and role-playing, showcasing its comprehensive performance [1].

Core Viewpoint - The article discusses the competition among AI large models in China, highlighting three different reports that claim different companies as the "first" in the industry based on varying criteria [2][30]. Group 1: Different "Firsts" Based on Metrics - The first report from Zhipu's prospectus claims that iFlytek is the first in terms of revenue, with over 500 million RMB projected for 2024 [6][8]. - The second report from IDC states that ByteDance's Doubao holds a 49.2% market share based on token usage, indicating it has the largest user base [10][12]. - The third report from A16Z and OpenRouter identifies DeepSeek as the global leader in open-source models, with 14.37 trillion tokens used, significantly higher than its competitors [15][18]. Group 2: Analysis of Business Strategies - iFlytek's ranking reflects its strong commercial capabilities and long-standing relationships in the government and enterprise sectors, focusing on B2B sales [19]. - ByteDance's Doubao leverages its massive user base from popular apps like Douyin and Toutiao, adopting a low-price strategy to capture market share [20]. - DeepSeek's appeal lies in its open-source nature and strong performance, particularly among global developers, although its monetization strategy is still under development [21]. Group 3: Implications for Job Seekers, Entrepreneurs, and Investors - For job seekers, the article emphasizes the importance of aligning personal skills with the respective business models of these companies [25]. - Entrepreneurs are advised to choose AI tools based on specific business needs rather than solely on market rankings [26]. - Investors should focus on the business models and monetization capabilities of these companies, as many are still struggling to achieve profitability despite high market shares [28]. Group 4: Future Directions in AI - The article suggests that the competition among these companies represents three distinct paths in the AI industry: the government-focused route (iFlytek), the user base-driven route (ByteDance), and the technology-driven route (DeepSeek) [27]. - It highlights the potential for multiple "firsts" to coexist in the market, similar to the current landscape of e-commerce in China [31]. - The article concludes that the ongoing competition signifies a significant shift in the global AI landscape, with Chinese companies gaining recognition and influence [33][34].

Core Viewpoint - The article emphasizes the importance of integrating artificial intelligence (AI) into various industries, particularly in the financial sector, to enhance service quality and operational efficiency while ensuring inclusivity and security in AI applications [3]. Group 1: AI Models - The choice between open-source and closed-source models is not just a technical issue but has profound implications for application. Open-source models promote equality and cost savings but may have slower iteration rates and higher error rates, while closed-source models offer stability and reliability but limit customization and transparency [4]. - The financial industry should focus on "AI+" rather than solely on building large models, combining the advantages of both open-source and closed-source models to enhance service quality and internal management efficiency [4]. Group 2: Decision-making AI vs. Generative AI - Decision-making AI excels in scenarios requiring high interpretability and accuracy, dominating over 80% of current applications in finance, particularly in risk assessment and fraud detection. In contrast, generative AI is more suited for creative tasks and is primarily used in non-core areas like customer service [5]. - The trend indicates that as the capabilities of large models improve, generative AI may see exponential growth and work in tandem with decision-making AI, blurring the lines between the two [5]. Group 3: AI Inclusivity and Computing Power - The demand for GPU computing power is expected to remain in a "tight balance" as AI becomes more widespread, necessitating efforts to optimize existing resources and expand capacity [8]. - Companies should adopt engineering methods to reduce operational costs and enhance resource efficiency while building high-performance computing centers to support AI applications [8]. Group 4: Safety and Security in AI Applications - As AI inclusivity increases, the stability and security of AI applications must be prioritized to protect public interests. This includes establishing safety measures and enhancing data quality to build trust in AI systems [9]. - There is a need to prevent model resonance to mitigate systemic risks, as the concentration of mainstream models may lead to vulnerabilities across institutions. Developing a reliable knowledge base and differentiated model training is essential for enhancing the resilience of the financial system [9].

Core Insights - The recent ranking of open-source AI models highlights the dominance of Chinese models, with DeepSeek, Qwen, Kimi, Zhipu, and MiniMax leading the global landscape, while OpenAI and Meta's models lag behind [3][5][25]. Group 1: Performance and Market Position - Chinese open-source models are rapidly closing the performance gap with closed-source giants, excelling in dimensions such as performance, pricing, ecosystem, and usability [5][25]. - Kimi's K2 Thinking model, featuring a trillion parameters, has outperformed OpenAI's GPT-5 and Anthropic's Claude 4.5 in various benchmarks [11][14]. - MiniMax M2 has also shown strong performance, ranking fifth in comprehensive lists, surpassing competitors like Gemini 2.5 Pro and Claude Opus 4.1 [14][79]. Group 2: Technological Advancements - The introduction of interleaved thinking in models like MiniMax M2 and Kimi K2 Thinking allows for more efficient task execution by alternating between action and reflection [34][36]. - MiniMax M2 employs a full attention mechanism, which, despite increasing training and inference demands, has proven to deliver better performance compared to sparse attention models [75][78]. Group 3: Cost and Accessibility - MiniMax's API offers competitive pricing at $0.3/$1.2 per million input/output tokens, although its verbose nature leads to high token usage, which can offset cost advantages [79]. - The open-source movement in China is gaining momentum, with MiniMax's release reinforcing the leadership established by DeepSeek and other Chinese AI labs in the open-source domain [80][84]. Group 4: Community and Developer Adoption - There is a growing recognition among developers for the practicality and affordability of Chinese open-source models, with many citing them as preferable alternatives to established closed-source options like OpenAI [25][84]. - The rapid updates and releases from various Chinese companies indicate a robust and collaborative open-source ecosystem that is continuously evolving [11][14].

责任编辑：栎树 财经频道更多独家策划、专家专栏，免费查阅>> 2025 小米人车家全生态合作伙伴大会于今日举行，罗福莉首次登场，并称"（小米开源模型的）代码能 力和agent能力，在世界级非常公开公正的评估榜单上，在我来看它已经进入了全球top1 2"。 ...

机器之心编辑部 英伟达的自研大模型，刚刚有了大版本的更新。 北京时间今天凌晨，英伟达发布了 Nemotron 3 系列开放模型，共三种规模，分别为 Nano、Super 和 Ultra ： 英伟达认为，随着企业从单一模型聊天机器人转向协同工作的多智能体 AI 系统，开发者正面临通信开销高、上下文漂移以及推理成本居高不下等挑战。同时，能 够支撑复杂工作流自动化的模型，必须具备足够的透明性与可解释性，才能赢得开发者与企业的信任。 其中 Nemotron 3 Nano 已在 Hugging Face 上线，是目前计算成本效率最高的模型，针对软件调试、内容摘要、AI 助手工作流和信息检索等任务进行了优化，可显 著降低推理成本。该模型采用独特的混合 MoE 架构，在效率与可扩展性方面实现了显著提升。 Nemotron 3 Nano 的总参数规模为 316 亿，激活参数规模为 32 亿（包含嵌入层为 36 亿）。在每次前向推理过程中，其激活的参数数量不到上代 Nemotron 2 Nano 的一半，却实现了更高的准确率。 与 Nemotron 2 Nano 相比，Nemotron 3 Nano 实现了最高 4 倍的 To ...

英伟达周一发布最新版开源人工智能模型系列Nemotron 3，并同步推出配套数据与工具，旨在为各行业提供透 明、高效、可定制的智能体AI开发能力。Nemotron3包含Nano、Super和Ultra三个版本，引入突破性的混合 潜在专家混合（latent MoE）架构，显著提升推理效率并降低运行成本。周一，英伟达股价开盘上涨近1.7%。 英伟达周一发布最新版系列开源人工智能模型"Nemotron"，以及配套的数据和库，旨在为各行各业提供透明、 高效、可定制的智能体AI（agentic AI）开发能力。该公司表示，这一新模型家族在速度、成本和智能水平方面 都将优于此前的产品。 Nemotron 3模型系列包括Nano、Super和Ultra三个版本，引入了一项突破性的混合潜在专家混合（latent Mixture-of-Experts，MoE）架构，帮助开发者以规模化方式构建和部署可靠的多智能体系统。 该公司表示，周一已经上线的Nemotron 3 Nano相比上一代产品效率更高，即运行成本更低，同时在处理包含多 个步骤的长任务时表现更好。另外两款体量更大的版本预计将在2026年上半年推出。 在Artifici ...

Market Overview - On December 15, US stock indices opened higher, with the Dow Jones up 0.33%, Nasdaq up 0.60%, and S&P 500 up 0.45% [1][2]. Key Stock Movements - The Nasdaq China Golden Dragon Index fell by 0.5%, with notable declines in popular Chinese stocks: Baidu down over 2%, Alibaba and Li Auto down over 1% [3]. - Tesla saw an increase of over 2%, as the company's VP confirmed the testing of a driverless ride-hailing service in Austin, Texas, after canceling the safety driver requirement [3]. - Nvidia rose by over 1% following the release of its Nemotron 3 series open-source model [3]. Company-Specific News - iRobot, a manufacturer of robotic vacuum cleaners, experienced a nearly 70% drop in stock price, triggering a trading halt. The company has filed for bankruptcy and will be acquired by its contract manufacturer, Picea Robotics [4][5].

风险提示及免责条款 市场有风险，投资需谨慎。本文不构成个人投资建议，也未考虑到个别用户特殊的投资目标、财务状况或需要。用户应考虑本文中的任何 意见、观点或结论是否符合其特定状况。据此投资，责任自负。 英伟达发布Nemotron 3开源模型系列。 ...