Huawei Cloud Announces Pangu Models 5.5 and All-new AI Cloud Service, Positioned as the AI Pioneer in Industries
Jun 20, 2025
[Dongguan, China, June 20, 2025] Huawei Developer Conference 2025 (HDC 2025) kicked off today at Dongguan Basketball Center. This conference includes an exciting lineup of inspiring keynotes, in-depth summits, insightful forums, and interactive workshops. Customers, partners, and Huawei speakers shared the latest technological innovations and achievements in HarmonyOS, Huawei Cloud AI Cloud Service, and Pangu models.
Zhang Ping'an, Executive Director of Huawei, CEO of Huawei Cloud, unveiled the next-generation Huawei Cloud AI Cloud Service. Built on CloudMatrix 384 supernodes, this service now delivers robust compute for advanced AI model applications. The launch was accompanied by the introduction of Pangu Models 5.5, featuring significant upgrades across five capabilities: natural language processing (NLP), computer vision (CV), multi-modal, prediction, and scientific computing. Mr. Zhang also showcased Pangu Models' transformative applications in agriculture, manufacturing, and scientific research sectors. Remaining steadfast in tackling industry challenges, Pangu Models continue to set new benchmarks as the AI pioneer in industries. Wang Yunhe, Director of Huawei Noah's Ark Laboratory, revealed the technical architecture behind Pangu Models. Zhang Yuxin, CTO of Huawei Cloud, then demonstrated how Huawei Cloud's full-stack AI innovations are transforming cloud services.
Launching the New-generation Huawei Cloud AI Cloud Service: The Best Compute for Foundation Model Applications
The unprecedented advancement of AI technologies is causing explosive growth in compute requirements for foundation model training and inference, leaving traditional computing architectures struggling to keep pace. Huawei Cloud's new-generation AI Cloud Service is based on CloudMatrix 384 supernodes. This supernode is the industry's first to implement peer-to-peer interconnection of 384 proprietary NPUs and 192 Kunpeng CPUs through a high-speed MatrixLink network to form a super AI server, which increases the inference throughput of a single card to 2,300 tokens/s, a near 4-fold improvement over that of non-supernodes. The supernode architecture can better support the inference of mixture of experts (MoE) models, with one expert per card. Furthermore, a single supernode can support the concurrent inference of 384 experts, greatly improving inference efficiency. The supernode also supports one operator task per card, and flexibly allocates resources, improves parallel task processing, reduces waiting time, and improves model FLOPS utilization (MFU) by more than 50%.
For model training tasks involving trillions or even tens of trillions of parameters, 432 supernodes can be cascaded into an ultra-large cluster with up to 160,000 cards in the cloud data center. Additionally, the supernodes support integrated deployment of compute for training and inference, performing tasks such as inference by day and training by night. The compute resources for training and inference can be flexibly allocated to help customers optimize resource usage.
Today, the Huawei Cloud AI Cloud Service has emerged as the preferred choice for AI infrastructure. It provides robust AI compute for more than 1,300 customers, such as Sina, SiliconFlow, ModelBest, Chinese Academy of Sciences (CAS), and 360, accelerating intelligent upgrade across industries.
Pangu Models 5.5: A Comprehensive Upgrade That Accelerates the Reshaping of Industries
Huawei Cloud's Pangu Models focus on industry-facing applications. Pangu Models help customers tackle the most challenging issues in their specific scenarios and reimagine both operations and efficiency across numerous industries. At this year's HDC, Huawei Cloud released Pangu Models 5.5, which have been fully upgraded to deliver new value for industries.
Huawei Cloud announces Pangu Models 5.5
Pangu Natural Language Processing (NLP) Model: The new 718B deep thinking model is a MoE model consisting of 256 experts. It has been drastically enhanced in knowledge reasoning, tool invoking, and mathematics, and it is among the industry's top-ranking models. Pangu Models are trained based on the full-stack software and hardware of the Huawei Cloud AI Cloud Service, which means the service can serve as the bedrock for building world-class large models.
Meanwhile, Pangu Models 5.5 have been upgraded to improve user experience in long-sequence processing, low hallucinations, integrated fast and slow thinking, and agents. For example, Pangu enables the technology of adaptive fast and slow thinking integration. By building difficulty-aware fast and slow thinking data and adopting two-phased progressive training, the model can adaptively switch between fast and slow thinking based on the difficulty of problems. In this way, the model can provide agile replies to simple problems and perform deep thinking on complex problems, improving the overall model inference efficiency by eight times. The Pangu DeepDiver uses key technologies such as long-chain problem synthesis and progressive rewards to achieve highly efficient execution in web page search and common Q&A. For example, DeepDiver can complete a more than 10-step complex Q&A within 5 minutes, and generate professional survey reports of over 10,000 words, greatly improving work efficiency for users.
Wang Yunhe, Director of Huawei Noah's Ark Laboratory
Pangu Models aim to help industry customers build their own models without "reinventing the wheel". Huawei Cloud provides enterprises with six core capabilities: Pangu foundation and industry-specific models, pre-training and post-training corpus, data engineering tool sets, model training tool sets, industry-specific judge models, and industry-specific evaluation platforms. With the complete toolchain and engineering methodology of Huawei Cloud ModelArts, enterprises can perform high-quality training, fine-tuning, and reinforcement learning on their accumulated data assets, in this way quickly building their own professional models.
The Chinese Academy of Agricultural Sciences (CAAS) has trained Pangu foundation models using large volumes of professional literature and cross-species multi-omics data in order to build its own Agricultural Scientific Discovery Model focused on scientific crop breeding. This model implements precise agricultural knowledge Q&A, efficient genetic analysis, and targeted site design to shorten the early R&D cycle and help researchers more accurately improve the target traits of crops. The CAAS team has already been successful in improving a rice strain based on the intelligent agricultural research system enabled by the Agricultural Scientific Discovery Model. This has led to the plant height being reduced by around 25% compared with conventional rice strains, while significantly improving lodging resistance and maintaining the same yield.
At the conference, Zhang Ping'an released five Pangu industry-specific deep thinking models covering the medical, finance, government, industrial, and automotive domains. These models will be officially launched in June, which will accelerate intelligent transformation across industries.
Pangu Multimodal Model: Huawei Cloud released a new model – the Pangu World Model based on the Pangu Multimodal Model. The Pangu World Model generates digital physical spaces for training intelligent driving and embodied AI robots, which also supports continuous optimization and iteration.
For example, in intelligent driving training, after the driving scenario, driving control information, and road network data of the first frame are input, the Pangu World Model can generate a driving video simulating that generated by a camera and a point cloud generated by lidar, along with a large amount of training data for intelligent driving without relying on costly real road video collection. Based on the Pangu Multimodal Model, Guangzhou Automobile Group (GAC Group) works with Huawei Cloud to achieve pixel-level mapping between videos (2D modality) and point clouds (3D modality). This means corner cases in complex scenarios can be reproduced within minutes, providing strong support for efficient end-to-end model iteration with one version iterated in just two days.
Meanwhile, Huawei Cloud has officially released the CloudRobo Embodied AI Platform based on the multimodal and thinking capabilities of Pangu Models. The platform integrates end-to-end capabilities, such as data synthesis, data labeling, model development, simulation verification, cloud-edge synergetic deployment, and sensing and security monitoring. The platform provides three core models to accelerate embodied AI innovation: embodied multimodal generation model, embodied planning model, and embodied execution model. The field of embodied AI poses many challenges, including a huge variety of robots, sensors, and interface protocols. Huawei Cloud proposes the Robot to Cloud (R2C) Protocol, and is hoping to work with robotics partners and industry organizations to build an open R2C protocol that will enable more robots to develop efficient and secure intelligence.
Pangu Prediction Model: This model uses the industry's first triplet transformer unified pre-training architecture, which realizes unified triplet encoding of data from different industries, including table data from manufacturing-process parameters, time series data from device-running logs, and image data from product inspections. The model efficiently processes and pre-trains this data within the same framework, greatly improving the accuracy of prediction and providing better generalization capabilities for predictions across different industries and scenarios.
Conch Cement uses the Pangu Model to predict the 3-day and 28-day strength of clinkers, providing scientific guidance for the preparation of raw materials. This allows Conch Cement to more flexibly reuse solid waste such as urban construction waste and industrial waste in the mixtures for raw materials, all while assuring high-quality cement. Conch Cement has thus been able to reduce costs, assist in urban waste disposal, and contribute to a greener environment.
Pangu Prediction Model has been applied to multiple industries, including steel manufacturing, non-ferrous metal manufacturing, and heating, to help industrial customers optimize processes and deliver optimal, system-level results. For example, since deploying the Blast Furnace Model, China Baowu Steel Group has been able to maintain a qualification rate of over 90% in molten iron temperature. This means that for each ton of molten iron, Baowu saves 2 kg of fuel, translating to a saving of 20 tons of fuel per blast furnace, per day. The Yunnan Company of Aluminum Corporation of China uses the Kun'an Model to predict the intervals for adding aluminum oxides, the amount of fluoride salt to be added, and the output of electrolytic cells. These new insights have enabled the company to improve the comprehensive production indicators of electrolytic cells, and reduce power consumption by 26 million kWh each year. Tianjin Energy Group has been using the Pangu Prediction Model to accurately predict future heating requirements. In the last heating season, Tianjin Energy Group was able to achieve 100% balanced heating and reduce energy consumption by 10%.
Pangu Scientific Computing Model: Huawei Cloud continuously deepens the combination of Pangu Scientific Computing Model and a wider range of scientific application fields. The Meteorological Bureau of Shenzhen Municipality further upgraded the Zhiji Model based on Pangu to implement regional ensemble forecast. Such forecast results more closely reflect changes in weather systems and help the Bureau forecast weather more accurately. The Chengdu-Chongqing region in China's Sichuan province has typically strong and intense rainfalls. In response, Chongqing Meteorological Service has built the Tianzi 12-hour Weather Forecast Model based on Pangu to enhance the capabilities of daily forecasting and warning against extreme weather. Shenzhen Energy Group uses Pangu to predict short- and mid-term wind and solar energy yields, which helps them adjust power generation more agilely to improve energy development efficiency.
Pangu CV Model: Huawei Cloud has released a 30B-parameter CV model based on the new MoE architecture. This is the largest CV model in the industry and supports multi-dimensional, pan-vision perception, analysis, and decision-making, with pan-vision meaning that it supports identification of images, infrared, lidar-generated point clouds, light spectrum, and radar. Furthermore, the Pangu CV Model uses a cross-dimensional generation model to create a pan-vision fault sample library. This library features rare-case fault samples in industrial scenarios such as oil and gas, transportation, and coal mining, greatly increasing the types of objects identified and improving the accuracy of identification in industry-specific scenarios.
CNPC has built the Kunlun Large Model based on Pangu and applied this model to more than 100 professional fields, such as exploration and development, oil refining and chemical engineering, and equipment manufacturing. In the equipment manufacturing field, the model is capable of detecting defects, such as porosity and tiny cracks in oil pipelines, with sub-millimeter precision. The model delivers about 40% higher identification efficiency and reduces manual workload by around 25%.
Over the past year, Pangu Models have been applied in more than 500 scenarios across over 30 industries. They have played a significant role in fields like government services, finance, manufacturing, healthcare, coal mining, steel, railways, autonomous driving, and meteorology, helping customers reshape industries.
At the conference, Huawei Cloud announced a brand upgrade. The announcement resonates with Huawei Cloud's commitment to pioneering AI compute and technology innovation, and helping every customer become a forerunner in the intelligent transformation of industries. Huawei Cloud is ready to work with customers, partners, and developers to accelerate intelligent transformation across industries.
Continuous Innovation: Reshaping Cloud Services with AI
Huawei Cloud drives continuous innovation in infrastructure and foundational models, transforming cloud services with AI and building an AI-native cloud that empowers businesses to fully harness AI and accelerate their intelligent journey.
At the event, Huawei Cloud CTO Zhang Yuxin unveiled ModelArts Versatile, the optimal AI agent platform for enterprises. ModelArts Versatile offers experience templates designed for diverse service needs, empowering businesses and developers to create professional, productive, and proactive enterprise-level AI agents. ModelArts Versatile also revolutionizes AI agent generation with its intelligent toolchain, transforming what once took days into mere minutes. By streamlining the process, it slashes both the complexity and expertise traditionally required for agent development.
Huawei Cloud has enhanced its intelligent assistant, Pangu Doer, leveraging cutting-edge AI compute, advanced Pangu models, and robust agents for unparalleled performance. The Pangu Deep Reasoning Model elevates Pangu Doer's intelligence by enhancing its intent understanding, task planning, and execution precision. Tailored professional domain models leverage specialized knowledge to boost expertise, while agentic workflows streamline customer's journey to address critical pain points. Task-driven toolsets are available on demand, optimized based on the extensive feedback from Huawei Cloud's customer service, ensuring Pangu Doer operates with unmatched accuracy and efficiency.
Zhang Yuxin, Huawei Cloud CTO, announced the fully upgraded Pangu Doer
CodeArts Doer, the intelligent assistant in CodeArts, Huawei Cloud's software development pipeline, is equipped with six specialized agents, which streamlines every stage of the R&D lifecycle—project management, product management, build, testing, and deployment—boosting efficiency by over 40%. Powered by advanced mechanisms like ArchRAG and motivated forgetting, it mimics human cognition, retaining valuable knowledge while discarding unwanted information, ensuring smarter, more accurate code generation. The full-link security has embedded 13 enterprise-grade code security standards, while self-healing workflows leverage multi-agent collaboration to detect and fix vulnerabilities in minutes.
Take the Sanxingdui Museum. Using CodeArts Doer, they crafted immersive digital experiences for visitors by building a digital app in just two days, transforming engagement effortlessly.
GaussDB Doer empowers enterprises to develop their own Database Administrator (DBA) capabilities with comprehensive upgrades across three key areas. For precise query, Huawei Cloud has built a professional O&M model based on the Pangu model, fine-tuned using over 10 billion tokens from tens of thousands of global GaussDB O&M cases. Dynamic knowledge graphs—built around fault symptoms, triggers, and diagnostic processes—enable continuous accumulation of O&M expertise for enterprises, enhancing issue query precision via Retrieval-Augmented Generation (RAG). To accelerate responses, GaussDB Doer embeds database troubleshooting workflows directly into its Chain of Thought (CoT) while offering specialized agents tailored to specific faults, seamlessly integrating over 50 troubleshooting tools for agentic AIOps. Furthermore, full compatibility with the Model Context Protocol (MCP) ensures flexibility, enabling users to choose from external tools for even greater efficiency in fault management.
MetaStudio revolutionizes virtual human creation with next-level ease and precision. Its advanced TTS voice synthesis delivers unmatched realism in tone and articulation, while more accurate lip sync and more dynamic gestures bring characters to life. Enhanced by 3A noise reduction and AI-powered voice activity detection (AI-VAD), interactions feel seamless, allowing natural interruptions just like real conversations. Powered by adaptable networking and intelligent routing algorithms, responses clock in under two seconds, ensuring smooth, engaging exchanges. Developers can effortlessly integrate these AI features into their projects with just five lines of code and a single SDK call.
The end-to-end model security solution safeguards every stage—from corpus and models to inference and applications. At the event, Huawei Cloud announced Model Application Firewall (MAF), designed to fortify model inference security by preventing prompt injections and detecting non-compliant content in real time. This system effectively counters common injections like jailbreaking, role playing, and malicious instructions. With a preconfigured library housing millions of prompt rules, it detects over 95% of prompt injections, boosting the overall model security score by more than 20%.
Building a New Ecosystem: A Better Choice for Global Developers
Huawei Cloud serves as a unified foundation for Huawei's foundational technologies including Huawei Cloud AI Cloud Service, HarmonyOS, Kunpeng, GaussDB, and EulerOS. More than 13 million developers have benefited from this integrated ecosystem, 8 million of which have signed up for Huawei Cloud.
To infuse AI into development, Huawei Developer Space has undergone comprehensive upgrades. This enhanced platform provides end-to-end resources from compute to applications, such as the AI-native application engine and HarmonyOS cloud phones. These upgrades offer developers instant access to Huawei Cloud compute, model, and agent resources.
Marking another milestone, the Huawei Developer Competition was launched today with a new track focused on proprietary AI full-stack technology. Global developers and technical leaders are welcome to participate in this challenge to push the boundaries of intelligent innovation.
Over the next two days, developers can participate in thought-provoking forums, roundtables, and hands-on activities such as CodeLabs, DTSE Challenges, and Pro Meetup, connecting with Huawei experts to spark infinite possibilities.