Summary
Huawei’s unveiling and open-source release of its groundbreaking Pangu AI models, including a 7-billion parameter dense model and a 72-billion parameter Mixture-of-Experts (MoE) model, marks a significant advancement in large-scale artificial intelligence development. These models are part of Huawei’s broader Ascend ecosystem strategy, aimed at accelerating AI research, industrial application, and innovation across diverse sectors by providing developers and enterprises with customizable, high-performance large language models (LLMs). The release includes not only model weights but also optimized inference code tailored for Huawei’s Ascend AI infrastructure, facilitating efficient deployment and fostering an open AI ecosystem.
The Pangu series employs a hierarchical architecture, comprising foundational large models specializing in natural language processing, computer vision, multimodal understanding, prediction, and scientific computing, alongside numerous industry-specific models trained on public datasets from sectors such as finance, government, and meteorology. The flagship Super Pangu model, featuring trillion-level parameters, supports complex cross-domain and multitasking scenarios, while innovations like Random Routed Experts (RRE) integrated with Transformer decoders enable significantly faster training throughput compared to conventional architectures. Trained over extensive multilingual and multi-domain datasets, these models demonstrate state-of-the-art performance in zero-shot and few-shot tasks including natural language understanding, open-domain dialogue, code generation, and machine translation.
Huawei’s open-source initiative reflects its commitment to transparency and collaborative AI development, allowing global developers and organizations to customize and improve these powerful models. This approach is complemented by Huawei’s full-stack AI ecosystem—encompassing hardware platforms like Kunpeng and Ascend chips, the MindSpore AI framework, and the ModelArts development pipeline—which together provide efficient, scalable AI training and inference capabilities. Practical applications already include AI-powered digital finance platforms and scientific domains such as weather prediction, illustrating the models’ versatility and industry impact.
However, Huawei’s AI advancements occur amid notable controversies and challenges. Security vulnerabilities inherent in early AI designs raise concerns about robustness against adversarial attacks, particularly in sensitive sectors. Additionally, Huawei has faced legal scrutiny, including investigations by Belgian authorities into alleged corruption linked to lobbying efforts in the European Parliament, leading to reputational and operational repercussions. Balancing rapid technological progress with ethical governance and security remains an ongoing challenge as Huawei continues to expand its influence in the global AI landscape.
Background
Huawei’s advancement in artificial intelligence has reached a new milestone with the development and open-sourcing of its Pangu AI models, including 7-billion and 72-billion parameter variants. This progress is deeply embedded within the broader context of the age of foundation models, where AI architectures are optimized to create synergy across devices, chips, and cloud infrastructures. This integration is exemplified by Huawei’s HarmonyOS, which incorporates AI technology at every layer—from the kernel to system applications—enabling what is termed as Harmony Intelligence. This framework fosters an open ecosystem that promotes collaboration while maintaining stringent privacy and security protections.
The open-source release of Pangu models is a strategic move by Huawei to implement its Ascend ecosystem strategy, accelerating the research, development, and industrial application of large model AI technologies. By making the source code freely available, Huawei empowers developers and enterprises to customize and adapt AI-powered large language models (LLMs) for diverse requirements, facilitating innovation across sectors.
Huawei has demonstrated the practical application of these technologies with the launch of its PanGu-powered AI Finance platform in December 2023. Introduced at the Huawei Cloud Fintech Summit, this platform aims to revolutionize the digital finance industry by integrating AI, big data analytics, and blockchain technologies to support Fintech firms globally.
The Pangu model architecture is organized into multiple layers to address various industry needs. The base layer (L0) comprises five fundamental large models, including natural language processing, visual, multimodal, prediction, and scientific computing models. The second layer (L1) features numerous industry-specific models trained on public data from sectors such as government, finance, manufacturing, mining, and weather forecasting. At the top-tier is the Super PanGu model with trillion-level parameters, designed for advanced scenarios requiring cross-domain or multitasking capabilities.
Huawei’s commitment to AI innovation is supported by its full-stack technology ecosystem, including the Kunpeng and Ascend AI compute cloud platform, the heterogeneous computing architecture CANN, the all-scenario AI framework MindSpore, and the AI development pipeline ModelArts. These technologies enable distributed parallel acceleration, operator and compilation optimization, and efficient cluster communication, resulting in AI model training speeds up to 1.1 times faster than popular GPUs. This comprehensive approach not only enhances Huawei’s core competitiveness but also helps customers and partners operationalize AI to create tangible industry value.
Together, these developments reflect Huawei’s broader mission to reshape industries through AI, leveraging breakthroughs in technology to meet evolving customer needs and drive innovation across multiple sectors.
Pangu AI Models
Huawei’s Pangu AI models represent a comprehensive family of advanced artificial intelligence architectures designed to address a wide range of applications across industries. Among them, the PanGu S Series includes the flagship Super PanGu model, which features trillion-level parameters tailored for complex AI scenarios such as cross-domain and multitasking applications. The architecture employs Random Routed Experts (RRE) combined with a Transformer decoder, enabling efficient extraction of sub-models suited for diverse tasks including conversation, translation, code generation, and natural language interpretation. This design achieves training throughput 6.3 times faster than comparable Mixture-of-Experts (MoE) models with similar hyperparameters.
The Pangu models adopt a hierarchical two-layer structure to optimize flexibility and industry adaptation. The first layer (L0) consists of five foundational large models specializing in natural language processing (NLP), visual understanding, multimodal tasks, prediction, and scientific computing. The second layer (L1) includes multiple large-scale industry-specific models trained on public datasets from sectors such as government, finance, manufacturing, mining, and meteorology. This decoupled architecture allows rapid customization and scaling, enabling customers to fine-tune models with their own data or selectively upgrade foundational or capability sets.
In 2024, Huawei open-sourced two major Pangu models: the dense Pangu 7B model and the hybrid expert Pangu Pro MoE 72B model, alongside an Ascend chip-based AI model integrated with advanced reasoning technology. The release includes the weights, basic inference code, and large-scale MoE inference code optimized for Huawei’s Ascend AI accelerators, supporting an open ecosystem for AI development.
The models are developed within Huawei’s MindSpore 5 AI framework and trained over more than 100 days on a cluster of 512 Ascend 910 AI chips, processing 329 billion tokens spanning over 40 natural and programming languages. This extensive training enables Pangu-Σ—an advanced iteration within the series—to outperform previous state-of-the-art models in the Chinese language domain across 16 zero-shot tasks, including few-shot natural language understanding, open-domain dialogue, question answering, machine translation, and code generation.
Huawei’s approach extends beyond pure model development by integrating a full-stack AI ecosystem. This includes the Kunpeng and Ascend compute platforms, the heterogeneous computing architecture CANN, the all-scenario MindSpore AI framework, and the ModelArts AI development pipeline. These tools collectively provide distributed parallel acceleration, operator optimization, and communication enhancements, allowing Pangu models to train up to 1.1 times faster than popular GPU alternatives. Additionally, Huawei Cloud offers industry-tailored Pangu models and development suites to facilitate rapid model training and deployment in sectors ranging from finance to meteorology.
The Pangu AI family also focuses on practical industrial adoption. For instance, Pangu Model 5.0 addresses large-scale industrial deployment challenges by supporting multiple NLP tasks like text generation, classification, and Q&A systems, thereby bridging the gap between foundational AI research and traditional industry requirements. The models’ scalability and modular design make them suitable for domain-specific applications, including knowledge-based question answering, coding assistance, banking agent systems, and safety risk detection.
Open-Source Initiative
Huawei’s open-source initiative for its Pangu AI models represents a strategic shift aimed at leveraging open-source technology to drive hardware sales and foster innovation in developing countries while challenging global competitors. By releasing the source code freely, Huawei enables developers and enterprises to customize and adapt the large language models (LLMs) according to their specific requirements, thereby accelerating AI application across various industries.
The company officially open-sourced the Pangu dense model with 7 billion parameters (Pangu 7B), as well as the Pangu Pro Mixture-of-Experts (MoE) model with 72 billion parameters (Pangu Pro MoE 72B), alongside the associated model inference code based on Huawei’s Ascend AI infrastructure platform. While the weights and inference code for the Pangu Pro MoE 72B model have already been made available, the Pangu 7B model’s weights and code were announced to be launched on the open-source platform imminently. Huawei has invited developers, corporate partners, and researchers globally to engage with these models, contribute feedback, and collaboratively improve the technology.
This initiative is a core part of Huawei’s broader Ascend ecosystem strategy, which aims to promote research, innovative development of large model technology, and accelerate AI-driven value creation across thousands of industries. The open-source approach not only facilitates widespread experimentation and deployment but also enhances the flexibility of the Pangu models by allowing industry-specific customization.
In addition to open-sourcing the models themselves, Huawei Cloud integrates Pangu with its development tools, such as CodeArts Snap—an intelligent programming assistant trained on vast datasets including 76 billion lines of code and 13 million technical documents. This integration exemplifies Huawei’s commitment to building an accessible and collaborative AI ecosystem that supports developers worldwide in operationalizing AI technologies and driving real value creation.
Furthermore, Huawei Cloud supports this initiative through comprehensive AI model toolkits, multi-scenario APIs, and a community platform offering quality courses and certifications. It also fosters collaboration via alliances that aggregate high-quality open data from various industry partners to build domain-specific datasets, thereby enhancing the quality and applicability of AI models. This ecosystem approach underlines Huawei’s vision of reshaping industries through AI by focusing on core technological competitiveness and partnership-driven innovation.
Huawei’s open-source Pangu models and their surrounding ecosystem illustrate a multi-layered AI platform strategy: from foundational models (L0) providing versatile capabilities, to large industry-tailored models (L1) trained on open datasets, and further to proprietary, scenario-specific models (L2) refined with customers’ own data—offering adaptable, intelligent solutions across diverse business needs. This layered architecture underscores Huawei’s commitment to making advanced AI technology accessible, customizable, and commercially valuable through open collaboration.
Training Data and Methodology
Huawei’s PanGu models, including the latest PanGu-Σ and versions with 7 billion and 72 billion parameters, have been trained on extensive and diverse datasets spanning over 40 domains. These domains include Chinese, English, bilingual corpora, and programming code, enabling the models to excel in tasks such as few-shot natural language understanding, open-domain discussion, question answering, machine translation, and code generation. The training data consists of 329 billion tokens sourced from more than 40 natural and programming languages, reflecting a broad multilingual and multi-domain approach.
The training methodology leverages Huawei’s proprietary MindSpore 5 framework, utilizing a cluster system equipped with 512 Ascend 910 AI accelerator chips. This infrastructure enabled PanGu-Σ to undergo training for over 100 days, emphasizing efficient processing and scalability in large-scale language model development. Architecturally, PanGu-Σ employs a Transformer decoder design integrated with Random Routed Experts (RRE), which allows dynamic sub-model extraction tailored for different application scenarios such as conversational AI, translation, code production, and natural language interpretation.
Huawei’s approach adopts a hierarchical and decoupled architecture consisting of two layers. The first layer (L0) includes five foundational large models specialized in natural language processing, vision, multimodal understanding, prediction, and scientific computing. The second layer (L1) comprises multiple industry-specific models trained on public data across sectors like government, finance, manufacturing, mining, and meteorology. This layered design allows rapid adaptation to diverse downstream tasks and facilitates customer-driven customization through independent dataset training and modular upgrades of foundation models or capability sets.
Moreover, models with over 10 billion parameters have demonstrated sufficient capability for domain-specific applications, including knowledge-based question answering, coding assistance, banking agents, and safety risk detection. To foster innovation and collaboration, Huawei plans to open-source model weights and inference code for PanGu 7B, encouraging global developers and researchers to contribute to the ecosystem, thereby accelerating research and practical AI applications across industries.
Model Reasoning and Inference Technologies
Huawei has developed advanced model reasoning and inference technologies to support its large-scale Pangu AI models. Central to this capability is the use of Ascend-based platforms that provide robust AI infrastructure, enabling efficient model training and inference. The Pangu Pro MoE (Mixture-of-Experts) model with 72 billion parameters and the Pangu 7 billion parameter dense model have both been open-sourced along with their inference code, facilitating wide accessibility and deployment.
The inference technology leverages heterogeneous computing architecture, known as CANN, and the MindSpore AI framework to optimize distributed parallel acceleration, operator compilation, and cluster communication. These optimizations result in training speeds up to 1.1 times faster than conventional GPU-based methods, significantly enhancing the performance of large models during reasoning tasks.
Huawei’s PanGu P Series also includes a professional version featuring a 10-billion parameter scale optimized for low-latency and cost-effective inference scenarios. Additionally, PanGu-Σ integrates Random Routed Experts (RRE) with a Transformer decoder architecture, enabling efficient extraction of sub-models for various applications such as conversational AI, translation, code generation, and natural language understanding. This architecture achieves a 6.3 times faster training throughput compared to traditional MoE models with similar hyperparameters.
The reasoning technology supports multi-modal data processing and natural language to API (NL2API) transformations, which improve semantic understanding and increase search accuracy by 15% in cloud search applications that handle videos, texts, and graphs. Huawei also combines its CodeArts development environment with Pangu models to offer CodeArts Snap, an intelligent programming assistant designed to enhance developer productivity.
By integrating AI computing power across devices, chips, and cloud platforms, Huawei enables real-time, on-demand model training and inference services via Ascend Cloud Service. This system supports models exceeding 10 billion parameters for domain-specific tasks such as natural language processing, computer vision, multi-modal understanding, knowledge-based question answering, coding assistance, banking agents, and safety risk detection. Furthermore, Huawei emphasizes systematic security measures to protect large model training and inference processes, ensuring trustworthy AI deployment across applications.
Applications and Use Cases
Huawei’s PanGu AI models, particularly the open-sourced 7 billion parameter dense model and the 72 billion parameter Pangu Pro Mixture-of-Experts (MoE) model, have been designed to address a wide range of industry-specific applications across multiple domains. The models support advanced natural language processing tasks such as text generation, understanding, translation, and code creation, enabling a broad spectrum of intelligent solutions.
Industry-Specific Models and Services
PanGu’s architecture is organized into three layers that facilitate tailored applications. The first layer (L0) includes foundational large models specialized in natural language processing (NLP), computer vision (CV), multimodal understanding, prediction, and scientific computing, serving as versatile building blocks for various scenarios. The second layer (L1) contains numerous large industry-specific models trained on public datasets from sectors such as government, finance, manufacturing, mining, and meteorology. These models provide ready-to-use services optimized for targeted business needs, helping customers deploy AI capabilities efficiently in their operations.
Huawei Cloud’s delivery of these models ensures consistency across different model sizes and allows integration with tools such as CodeArts Snap, an intelligent programming assistant that enhances developer productivity through AI-powered coding support. Moreover, multi-round natural language interactions and multi-modal embeddings improve semantic understanding, enhancing the accuracy of cloud search applications by approximately 15%, which is particularly useful in handling videos, texts, and graph data.
Advanced Capabilities and Scenario Adaptability
The latest PanGu Model 5.0 offers three core features—adaptability to diverse business scenarios, multi-style modeling, and advanced intelligence—that enable customized AI solutions across
Performance and Benchmarks
Huawei’s PanGu series of AI models demonstrate significant advancements in large-scale language model performance across multiple domains. The PanGu P Series includes a professional version featuring a 10-billion parameter scale, optimized for scenarios requiring low-latency and cost-efficient reasoning. The PanGu-Σ model incorporates innovative Random Routed Experts (RRE) within a Transformer decoder architecture, enabling flexible extraction of sub-models tailored for applications such as conversation, translation, code generation, and natural language understanding. This design achieves a training throughput approximately 6.3 times faster than comparable Mixture-of-Experts (MoE) models with equivalent hyper-parameters.
In the Chinese language domain, PanGu-Σ surpasses previous state-of-the-art models in zero-shot evaluations across 16 distinct tasks. Its training spans datasets from 40 diverse fields, including Chinese, English, bilingual corpora, and code, enabling strong performance in few-shot natural language understanding, open-domain dialogue, question answering, machine translation, and code production. Models with parameter counts exceeding 10 billion have demonstrated sufficient capacity for various domain-specific tasks such as natural language processing, computer vision, and multi-modal applications like knowledge base question answering, coding assistance, banking agents, and safety risk detection.
Huawei’s open-source release includes the PanGu dense model with 7 billion parameters and the PanGu Pro MoE model with 72 billion parameters, alongside inference technologies based on the Ascend platform. The availability of these models and related inference code facilitates large-scale AI infrastructure deployment and experimentation.
Furthermore, Huawei offers a suite of capability sets consistent across model sizes, encompassing knowledge-based question answering, copywriting, code generation, as well as image generation and understanding for multimodal models. The L1 layer of their AI ecosystem comprises industry-tailored models trained on open industry datasets targeting sectors such as government, finance, manufacturing, mining, and meteorology, thus ensuring practical applicability and enhanced domain performance.
Collectively, these performance characteristics position Huawei’s PanGu models as competitive solutions in both research and industrial settings, delivering efficient training, robust zero-shot and few-shot generalization, and comprehensive domain versatility.
Security, Ethics, and Responsible AI
The security of AI systems is a critical concern, particularly regarding the protection of model integrity and data confidentiality. Vulnerabilities stemming from the lack of explainability in AI models open doors for adversarial machine learning attacks such as evasion, poisoning, and backdoor attacks. These attacks are notably effective and exhibit strong transferability across different machine learning models, posing significant threats to deep neural network (DNN)-based AI applications. Because security considerations were often overlooked during the initial development of AI algorithms, attackers can manipulate inference results, potentially causing serious misjudgments. This risk is especially severe in sensitive domains like healthcare, transportation, and surveillance, where successful attacks could lead to property damage or endanger human safety.
In the context of responsible AI development, transparency and ethical governance are paramount. However, Huawei’s AI initiatives have faced scrutiny beyond technical challenges. Belgian federal prosecutors have launched investigations into alleged active corruption within the European Parliament linked to Huawei, with accusations suggesting bribery intended to secure favorable lobbying outcomes. This controversy echoes the 2022 Qatargate scandal and has intensified concerns regarding lobbying practices in the European Union. In response to these allegations, the lobby organization SolarPower Europe removed Huawei from its membership in May 2025.
To foster responsible AI innovation, Huawei has committed to open-source practices with its Pangu AI models. Open sourcing the original source code allows developers and enterprises to access, test, and customize AI-powered large language models (LLMs), promoting transparency, collaborative improvement, and ethical use. This approach not only advances AI capabilities but also encourages the development of secure and accountable AI systems by enabling thorough inspection and modification of model architectures and reasoning technologies.
Impact on AI Research and Industry
Huawei’s unveiling of the open-source 7B and 72B Pangu AI models, alongside their cutting-edge model reasoning technology, marks a significant milestone in advancing both AI research and industrial applications. This development aligns with China’s 2024 AI+ action plan, which positions artificial intelligence as a crucial engine for fostering new productive forces across various sectors.
In the research domain, Huawei’s full-stack innovation—spanning hardware, software, and development frameworks—enhances the efficiency and scalability of AI model training. Leveraging a heterogeneous computing architecture (CANN), an all-scenario AI framework (MindSpore), and an integrated AI development pipeline (ModelArts), Huawei enables distributed parallel acceleration and optimized compilation techniques. These advancements allow training speeds up to 1.1 times faster than popular GPUs, significantly accelerating AI research and experimentation.
From an industrial perspective, Huawei’s AI technologies are already widely applied in traditionally challenging sectors such as coal mining, smelting, oil and gas, and the chemical industry. The company’s establishment of the Oil, Gas & Mining Business Unit in 2021 demonstrates a focused effort to harness AI for manufacturing industries, despite facing key challenges that hinder smooth AI integration. The open-source release of large-scale models and enhanced reasoning capabilities is expected to lower barriers to AI adoption, facilitate customized industrial solutions, and drive smarter, more efficient production processes.
Moreover, Huawei’s commitment to building a robust AI chip ecosystem ensures that their AI innovations remain resilient amid geopolitical challenges, such as U.S. restrictions. This ecosystem supports sustainable AI development and deployment across multiple domains, reinforcing Huawei’s role as a major contributor to the global AI landscape. Together, these advances position Huawei’s Pangu models and technologies as catalysts for accelerating AI-driven transformation in both research and industry.
Criticisms and Challenges
Huawei’s advancements in AI, including the open-source release of the 7B and 72B Pangu models, have not been without criticisms and challenges. One major concern centers on the security vulnerabilities inherent in early AI development. Due to insufficient security considerations during the initial design of AI algorithms, attackers have exploited these weaknesses to manipulate inference results, potentially causing misjudgments with serious consequences. This issue is particularly critical in sectors such as healthcare, transportation, and surveillance, where security breaches can lead to property loss or even threaten personal safety.
Beyond technical vulnerabilities, Huawei faces reputational and regulatory challenges. In May 2025, Belgian federal prosecutors launched an investigation into allegations of active corruption within the European Parliament, with reports suggesting that Huawei was the intended beneficiary of the alleged bribery. This scandal drew parallels to the 2022 Qatargate incident and raised heightened concerns about lobbying practices within the European Union. The fallout from this investigation led to the removal of Huawei from the lobby organization SolarPower Europe, signaling potential setbacks in the company’s influence and collaboration efforts within the European market.
From a technical standpoint, while large-scale models such as those in the Pangu series—ranging from 7 billion to 72 billion parameters—have demonstrated impressive capabilities across various domains like natural language processing, computer vision, and multimodal tasks, deploying these models effectively remains challenging. Balancing model scale with latency and cost, particularly for industry-specific applications, requires careful engineering and resource optimization. Huawei’s approach, exemplified by their layered model architecture and simulation engines for resource management, addresses some of these issues but does not fully eliminate the complexities inherent in scaling AI technologies for practical use.
Together, these criticisms and challenges illustrate the complex landscape Huawei navigates as it pushes forward with its AI innovations. While the company continues to drive science and technology development aligned with customer needs, it must also address security vulnerabilities and navigate legal and ethical scrutiny to sustain its technological leadership and market position.
Future Directions
Huawei’s future directions in AI development focus on leveraging both scientific research and customer-centric innovation to drive the next wave of technological advancements. In 2024, the company intensified its commitment to foundational research and theoretical studies, with an emphasis on mathematical logic to better align AI capabilities with evolving business strategies.
A significant part of Huawei’s forward-looking strategy involves the open-source release of its Pangu 7B model weights and inference code. By inviting developers, corporate partners, and researchers worldwide to access and customize these resources, Huawei aims to foster a collaborative ecosystem that accelerates innovation and the practical application of AI technologies across diverse industries. This open-source approach is integral to the company’s Ascend ecosystem strategy, which seeks to promote large model technology research and facilitate value creation in thousands of sectors.
Furthermore, Huawei is actively engaging in ecosystem building by forming high-quality data alliances, such as the partnership between Huawei Cloud and various industry committees. These alliances aim to aggregate open datasets and create industry-specific training data, thereby enhancing the quality and performance of AI models. Huawei recognizes developers as a crucial driving force behind digital innovation and is positioning its ecosystem programs to empower this community.
In alignment with China’s broader AI+ action plan that identifies AI as a key driver for new productive forces, Huawei continues to apply AI across traditional industries like coal mining, smelting, oil and gas, and chemical manufacturing. The company is addressing key industrial challenges through AI-enabled solutions, particularly focusing on enhancing efficiency and innovation in manufacturing sectors. Through these integrated efforts, Huawei envisions expanding the impact of its AI models, such as Pangu, to transform industries and create sustainable technological growth globally.
The content is provided by Blake Sterling, 11 Minute Read
