Weekly Enterprise Watch: Meituan Releases LongCat-2.5 Trillion-Parameter Model, Robots Take to Their Posts en Masse
Meituan released its native multimodal large model LongCat-2.5, with 1.6 trillion total parameters, trained entirely on domestic compute—the largest training task completed on domestic compute to date; its scores on multiple mainstream benchmarks fall between Claude-Opus-4.8 and Claude-Opus-5, with some benchmarks surpassing Claude-Opus-5. Moonshot AI released Kimi Hosted Agents, offering the Agent runtime framework as a cloud service with a 7×24 hosted environment. AgiBot, together with Tsinghua University and Infinigence AI, released the collective experience learning paradigm CEL; in a phone packaging task, the number of robots expanded from 16 to 96, and the success rate rose from 86.7% to 98.3%.
Body · AI translation
This week, the yardstick for AI shifted from parameter size to success rate. The first seminar of Dahua Technology's European Compliance Special Working Group was held in the Netherlands. Recently, the first seminar of Dahua Technology's European Compliance Special Working Group was held in the Netherlands. Relevant leaders and compliance experts from various European regions, headquarters and many locations around the world, as well as external senior lawyers, participated together to further deepen the localized practice of global compliance governance. Gosuncn: Controlled subsidiary Chuanglian Technology terminates its Beijing Stock Exchange listing tutoring filing. Gosuncn announced that its controlled subsidiary Chuanglian Technology, based on consideration of future strategic planning, after friendly consultation with Shenwan Hongyuan, both parties agreed to terminate the Beijing Stock Exchange listing tutoring agreement and signed a termination agreement on November 4, 2026. Shenwan Hongyuan withdrew the listing tutoring filing application with the Zhejiang Securities Regulatory Bureau on the same day. This matter does not require review by the company's board of directors or shareholders' meeting and will not have a significant impact on production and operations. Unitree Technology's Wang Xingxing: The core challenge of physical AI lies in the alignment between models and the physical world. At the 9th Hongqiao International Economic Forum, Wang Xingxing, founder and chairman of Unitree Technology, appeared at the sub-forum on "Innovative Development Cooperation for Humanoid Robots and Embodied Intelligence Industry." Speaking of the core challenges of current physical AI, Wang Xingxing believes they mainly lie in the alignment between models and the physical world. In his view, the understanding capability of multimodal models themselves is already relatively good; the biggest problem is the error in input-output alignment, including deviations at the input level between sensor measurements and the true state, and at the output level between action commands and actual motion. Wang Xingxing judges that this part of the problem may be solved in the next 1-2 years. Seagate helps AI and cloud infrastructure efficiently scale high-capacity storage. The Seagate Mozaic™ hard drive uses heat-assisted magnetic recording (HAMR) technology and, combined with a shingled magnetic recording (SMR) technology configuration, has become the first and currently only SMR hard drive included in Broadcom's host bus adapter (HBA) certified supplier list. HBAs are used to connect servers and storage devices to support large-scale data center deployment and hard drive management. Lingguang 2.0 manages over 1.16 million affairs in its first week online, with users exceeding 20 million. November 5 news: Ant Group's AI assistant Lingguang announced the latest data: since Lingguang 2.0 went online on October 27 this year, within one week it has helped users manage over 1.16 million affairs and executed over 3.6 million tasks. Currently, Lingguang users have exceeded 20 million. Tencent WorkRally features comprehensively upgraded. Recently, Tencent's AI content creation platform WorkRally completed a new round of upgrades, expanding its production capabilities for mass creators from comic dramas to short dramas, advertisements, creative videos and other types, further lowering the threshold for use by OPC (One-Person Company) and individual creators. Through collaboration across scriptwriting, shot generation, local editing and post-production, it helps creators reduce trial and error and rework while polishing work quality, optimizing the overall production cost from conception to finished film. Qiyuan Robotics completes its first OTA version upgrade, with Qiyuan Q1 and T1 simultaneously receiving capability iterations. Qiyuan Robotics officially pushed an OTA version update, and the two personal robots Qiyuan Q1 and Qiyuan T1 simultaneously received their first capability iteration and upgrade since launch. Among them, Qiyuan Q1 opened the two major creation platforms PrimeTOOLKIT and PrimeFLOW; in addition, Qiyuan T1 unlocked stair recognition automatic transformation passage and audio-video call remote control capabilities; both products also simultaneously launched educational functions such as AI enlightenment and spoken English practice. Moonshot AI: Kimi Hosted Agents officially released. Moonshot AI announced the release of Kimi Hosted Agents (hosted agents), providing the Harness that supports Agent operation as a cloud service to enterprises and developers. Enterprises can obtain a 7×24-hour hosted environment without building and maintaining the operating framework themselves, supporting long-horizon task execution and efficient multi-agent collaboration. UBTECH's Thinker-Sim embodied intelligence simulation platform released. According to UBTECH, the UBTECH Thinker-Sim embodied intelligence simulation platform was officially released. According to the introduction, UBTECH has built Thinker-Sim, an embodied intelligence simulation platform centered on Real2Sim2Real, constructing a one-stop technical system covering digitization of real scenarios, large-scale simulation data synthesis, and robot capability evaluation. Meituan releases the LongCat-2.5 large model. On November 2, Meituan released a new-generation large model LongCat-2.5, a native multimodal model oriented toward long-horizon agent tasks, with total parameters of 1.6 trillion. LongCat-2.5 continues the R&D direction of the previous LongCat series models and was trained entirely on domestic computing power, making it the largest training task completed on domestic computing power to date. Model evaluation results show that in multiple mainstream benchmark evaluations, LongCat-2.5's scores fall between Claude-Opus-4.8 and Claude-Opus-5, and in some evaluations it surpasses Claude-Opus-5. Sugon: Sugon storage systems have been applied in multiple scenarios including Sugon 8000, robot data collection and training, and autonomous driving. On October 29, at the 2026 Data Storage Industry Conference held in Zhengzhou, Sugon showcased its latest exploration for the Token era, using the "AI Data Factory" to drive storage to actively participate in the entire AI production process and improve Token output efficiency. The "AI Data Factory" focuses on connecting all stages of data use, enabling AI to obtain the information it needs in a timely manner and reuse existing computation results, thereby improving computing power utilization efficiency. Currently, Sugon storage systems have been applied in multiple scenarios including Sugon 8000, robot data collection and training, and autonomous driving. Tencent Maps releases LBSwork and unveils the full Agent capability landscape, building an Agent-friendly map. On October 30, the 2026 Tencent Global Digital Ecosystem Summit was held in Shenzhen. At the Tencent Maps industry application session, Tencent Maps officially unveiled the full Agent capability landscape, fully opening up map Agent capabilities to industry customers, smart hardware, and developers. At the same time, it released LBSwork, a desktop AI map workbench, which, based on the WorkBuddy same-origin Harness framework, provides one-stop "AI + map" capabilities for customers in retail, travel, cultural tourism, and other industries. AgiBot, together with Tsinghua University and Infinigence AI, releases the CEL collective experience learning paradigm, achieving autonomous cluster evolution with 96 real robots. On October 30, AgiBot, together with Tsinghua University and Infinigence AI, released the collective experience learning paradigm CEL. This paradigm uses the real operation trajectories of robot companions as the source of experience, and is paired with the FERT failure-guided experience retrieval fine-tuning algorithm and the FATS fault-tolerant asynchronous training system. On five real tasks—flower arrangement, clothes folding, phone charging, stool assembly, and phone packaging—the policies trained with CEL all exceeded the average level of 6 human teleoperators, leading by up to 16.7 percentage points. In the phone packaging task, the number of robots participating in training expanded from 16 to 96, and the success rate rose from 86.7% to 98.3%. D-Robotics and Unitree Robotics jointly launch an embodied intelligence education solution. At the Changsha Education Equipment Exhibition, D-Robotics and Unitree Robotics appeared together in a joint exhibition format for the first time and announced a deep cooperation. With the Sunrise S100 computing power box as the core computing power foundation, the two parties combined Unitree Robotics' Go2 quadruped robot and R1 humanoid robot to jointly launch an embodied intelligence education solution, covering hardware equipment, curriculum systems, course-competition integrated practice projects, and teaching environment construction, providing complete support for primary and secondary schools to systematically carry out embodied intelligence education. The two parties will take education as the first scenario for cooperation to be implemented, jointly create a replicable and promotable regional robot education demonstration model, and plan to further expand cooperation to more scenarios such as industry. Apple and Alibaba reach a cooperation agreement, integrating Qwen into Apple Intelligence. On October 31, Apple and Alibaba announced a cooperation to integrate Qwen into Apple Intelligence. Chinese users will be able to directly invoke Qwen's text and image understanding, content generation, and other capabilities in Siri, writing tools, and visual intelligence. These features will open for testing with the latest iOS developer Beta starting November 4, and will be launched for Chinese users when iOS 27.3 is officially released in December.