Core Insights - The article highlights the rapid development and release of next-generation AI models by Beijing-based companies, positioning Beijing as a leader in the global AI landscape [1][2][15] Group 1: Model Releases and Performance - In early 2026, several AI companies in Beijing, including Douyin and Zhiyuan AI, launched new models, achieving significant breakthroughs in various AI fields such as general language models and multi-modal video generation [2][3] - Zhiyuan AI's GLM-5 model ranked fourth globally and first among open-source models, while Douyin's Seedance 2.0 received acclaim for its advanced video generation capabilities [2][4] - Kimi K2.5 achieved the best performance in multiple agent evaluations, showcasing the competitive edge of Beijing's AI models [2][3] Group 2: Advancements in Capabilities - The evolution from "writing code" to "completing engineering projects" is exemplified by GLM-5, which allows developers to accomplish tasks that traditionally required months of work in a matter of hours [3][4] - Kimi K2.5 introduced an agent cluster system, enabling collaborative task execution and significantly enhancing efficiency in complex problem-solving [5][6] Group 3: Multi-Modal and Content Production - Seedance 2.0 marks a shift from AI as a creative tool to a reliable production tool, enabling high-quality video generation with reduced costs and increased efficiency [7][8] - The model's capabilities allow for seamless integration of audio and visual elements, enhancing the overall content creation process [7][8] Group 4: Embodied Intelligence and Robotics - Companies like Galaxy General are pioneering embodied intelligence, integrating advanced robotics into various sectors, including retail and manufacturing [9][10] - The launch of the Galbot S1 robot signifies a leap in industrial applications, showcasing autonomous capabilities in real-world scenarios [9][10] Group 5: Open Source and Collaboration - Open-source strategies have become crucial for Beijing AI companies, allowing them to compete effectively with international counterparts while maintaining high efficiency with limited resources [13][14] - The collaboration between AI firms and domestic chip manufacturers has established a robust ecosystem for the development of AI models [12][14] Group 6: Financial and Market Environment - Beijing has fostered a supportive capital market for AI development, with long-term investment strategies that encourage innovation and reduce risk for emerging companies [14] - The successful IPOs of leading AI firms in Beijing reflect the growing confidence and investment in the AI sector [14] Group 7: Global Impact and Future Outlook - The advancements in AI models from Beijing are gaining international recognition, with global developers eager to engage with these technologies [15][17] - The article emphasizes the transition of AI tools from being exclusive to a select few to becoming accessible production tools for a broader audience, marking a significant milestone in AI development [17]

金磊 发自 凹非寺 量子位 | 公众号 QbitAI 这就是由 生数科技 联合 清华大学 ，正式开源的大一统世界模型—— Motus 。 项目主要负责人，是来自清华大学计算机系朱军教授TSAIL实验室的二年级硕士生 毕弘喆 和三年级博士生 谭恒楷 。 之所以说是大一统，是因为Motus在架构上，直接把VLA（视觉-语言-动作）、世界模型、视频生成、逆动力学、视频-动作联合预测这五种具 身智能范式， 首次 实现了"看-想-动"的完美闭环。 而且在50项通用任务的测试中，Motus的绝对成功率比国际顶尖的 Pi-0.5 提升了 35% 以上，最高提升幅度甚至达到了 40%！ 在Motus的加持之下，现在的机器人已经具备了 预测未来 的能力。 国产开源 具身世界模型 ，直接秒了Pi-0.5，而且还是几位 清华硕、博士研究生 领衔推出的。 瞧， Cloudflare人机验证 任务，机器人可以轻松拿捏： 从视频中不难看出，面对形状不规则的曲面鼠标，Motus控制的机械臂不仅能精准识别，还能根据鼠标与屏幕点击框的距离，平稳连续地移 动，最后极度精准地完成点击。 再如长程多步推理的 孔明棋 任务，Motus同样展现出了严密 ...