混乱特工

China’s first supernode for 2t-parameter models enters service in Ulanqab_我的网站

兵临城下

A | 智能载体如何更好地像人类一样拥有理解物理世界的能力？这是当前AI落地仍面临的一大关键问题。依托“物理世界数字化”的入口之一，即大量的物联感知设备，全球智能物联龙头海康威视（002415.SZ）在不断探索破解这一问题。在2026世界人工智能大会（WAIC 2026）期间，《海康观澜大模型白皮书（2026版）》正式发布。过去几年，海康威视一直在AI大模型上持续投入，观澜大模型则正式发布于2023年。

B |

Alibaba Cloud's Lingjun Zhenwu M890 supernode instance Photo: Courtesy of Alibaba Cloud

Lingjun Zhenwu M890 supernode instance Photo: Courtesy of Alibaba CloudAlibaba Cloud on Tuesday officially launched its Lingjun Zhenwu M890 supernode instance in Ulanqab, North China's Inner Mongolia Autonomous Region, with the first batch of instances now available for sale in the region.
The instance is designed to handle inference for mixture-of-experts models with up to 10 trillion parameters, the company said in a statement sent to the Global Times on Wednesday.
This marks the first supernode-form computing architecture in China to successfully run large language models exceeding 2 trillion parameters, according to the company.
Industry expert Tian Feng told the Global Times that the commercial rollout of supernode infrastructure could significantly reduce training cycles, lower costs, and speed up iteration for AI developers requiring massive computational resources.
The company said the new instance has already been used to power commercial services for large language models such as KimiK3 and Qwen3.8Max.
The Lingjun Zhenwu M890 supernode instance supports FP8/FP4 low-precision computing. Through the ICNSwitch 1.0 chip, its scale-up interconnect scale has been expanded from 16 cards to 64 cards, with inter-card interconnect bandwidth boosted to 800 GB/s. Enterprises can provision 64-card, high-speed-interconnect computing units through the cloud without building their own data centers, according to the company.
In training scenarios such as autonomous driving and embodied intelligence, the instance delivers three times the training performance compared with the previous-generation Zhenwu 810E, the company said.
Ulanqab, where the supernode instance debuted, is one of Alibaba Cloud's five super data centers. The facility sources approximately 90 percent of its electricity from green energy, providing a low-carbon operating environment for high-density computing power.
Leveraging its climate, energy and network advantages, Ulanqab has transformed from "China's potato hometown" into the "token factory" - a term increasingly used in the AI industry to describe infrastructure dedicated to producing the digital building blocks generated by large language models.
By the end of 2025, the city had attracted 84 data center projects, including 81 intelligent computing centers, with total investment exceeding 500 billion yuan ($74.1 billion) and operational computing power reaching approximately 172,000 petaflops, ranking it firmly in the nation's top tier, according to domestic media reports.
On August 6, China's largest AI computing industrial park was completed and put into operation in Ulanqab. The project highlights a broader race in China to build massive AI data centers capable of supporting the next generation of AI models while addressing soaring electricity demand, according to Chinese experts.
In recent years, Inner Mongolia has been rapidly positioning itself as a global-scale AI computing center cluster. Major technology companies, including Huawei, Tencent, ByteDance and Alibaba; telecom operators China Mobile, China Telecom and China Unicom; as well as cyberspace infrastructure service provider VNET, have established computing facilities in the region.
As the AI industry gradually transitions from the training era to the inference era and large model parameters continue to expand, supernodes have become a central battleground for AI infrastructure.
Chinese vendors are accelerating deployments in this space. Huawei has commercially deployed more than 750 sets of its Ascend 384 supernodes across industries including internet, telecom operators, finance, education, healthcare, transportation and manufacturing. It is also the only domestic supernode to have trained state-of-the-art (SOTA) models.
Baidu AI Cloud has also launched its Tianchi 256-card supernode based on Kunlun chips, with support for major models including Wenxin, DeepSeek, GLM, and MiniMax.
Meanwhile, supercomputer manufacturer Sugon has unveiled China's first fully domestic 100,000-card AI supercluster Sugon 8000 (Dengfeng), integrating supercomputing and AI computing on a unified architecture. It has now been connected to the national supercomputing internet to provide computing services to government, research, and enterprise clients nationwide.
Tian, former dean of SenseTime's Intelligence Industry Research Institute, told the Global Times that the flurry of domestic supernode launches reflects a broader inflection point as China's AI sector pivots from training capacity toward efficient, large-scale inference.
The expert further said that the commercial viability of these systems - evidenced by Huawei's extensive deployed base, Baidu's rapid model adaptation and Sugon's integration into the national computing network - suggests domestic vendors are moving beyond proof-of-concept to genuine production-grade infrastructure, a prerequisite for sustaining the next wave of trillion-parameter model proliferation, Tian noted.
The move also underscores China's push for self-reliance in AI infrastructure as US chip export restrictions continue to tighten, Tian said, noting that, in the supernode domain, Chinese companies are shifting from imported graphics processing units toward homegrown interconnect chips and domestic compute clusters, a transition that could reshape value allocation across the AI industry chain.
。作为源自物理世界的大模型，海康观澜大模型致力于更好地理解真实世界，可处理视频、图像、语言、音频、X光、毫米波等丰富的物联感知信息，并针对行业应用进行垂类适配和调优，服务千行百业。

C | 据最新介绍，海康观澜大模型技术体系覆盖“基础大模型-垂类大模型-大模型产品-行业应用”，具备“多快准省”的系统性优势。

D | 例如，其中的“多”即体现在对物理世界的全面感知和对行业需求的广泛覆盖：在模态方面，构建了视觉大模型、X光大模型、毫米波大模型、语言大模型等完整的物联感知能力矩阵；产品方面，已发布上千款大模型软硬件产品，覆盖云边端全域；场景方面，深耕90多个垂直行业、2000多个场景，沉淀了600多种智能化解决方案。

E | 又例如“准”，海康威视称，这主要来自信号质量、模型精度与软硬件系统的协同提升。

F | 截至目前，周界防范误报率降低90%以上，工业微孔缺陷检出率达99.99%，安检场景下X光安检检出率达97%以上。

G | 值得关注的是，海康威视还着重提出，垂类大模型更“懂行”，可高性价比适配各类场景。基于基础大模型，海康威视面向特定领域深度注入行业知识和业务Know-how，已形成了覆盖安全生产、工业制造、连锁巡检、园区、公共安全、城市治理、交通管理、自然灾害监测、基础设施巡检等领域的系列垂类大模型。相比通用大模型，垂类大模型训练推理成本更低，可快速、高性价比地适配各类场景。以工业制造领域目前落地情况来看，海康威视的工业大模型覆盖产线作业SOP检测、工业质检、工业测温等核心环节，助力提质、增效、降本、安全。例如，通过海康睿影X光缺陷检测系列产品，深度服务于电子、新能源、半导体、汽车、食药品等多个行业，在电子缺陷检测领域，可以微米级自动识别电路板焊接气泡缺陷，检出率高，误判率＜1%，适应多种缺陷检测场景；工业测温大模型则显著提升极端工况下的测温精度，有效抑制强水汽、风沙等环境干扰，确保3200℃高温工艺监测“看得清、测得准”；同时支持最小35μm微小目标的精准捕捉，并将远距离测温能力从传统50米拓展至300米，突破距离瓶颈，为工艺生产、产品研发及新能源发电等场景提供了更可靠、更精细的温度采集保障。 AI深入物理世界的实践路径仍在不断摸索。海康威视方面称，这些创新的不断落地，标志着AI正逐步嵌入社会、产业当中，真正助力千行百业在智能化转型的深水区找到升级路径。