Summary

Huawei’s Pangu Pro MoE (Mixture of Experts) model is a notable large language model (LLM) developed by Huawei’s AI research division, Noah’s Ark Lab, showcasing significant advancements in scalable AI architecture and efficiency. Built on Huawei’s proprietary Ascend AI chips, the Pangu Pro MoE 72-billion parameter model employs a Mixture of Grouped Experts (MoGE) design to optimize workload balancing and inference performance, representing a strategic effort to advance China’s AI capabilities and compete with global technology leaders. The model is part of Huawei’s broader Pangu AI series, which includes dense and hybrid architectures released under open-source licenses to foster community engagement and industry adoption.
The model gained prominence not only for its technical innovations but also due to controversy arising shortly after its open-source release in mid-2023. A group named HonestAGI alleged that Pangu Pro MoE was not an original creation but rather an “upcycled” derivative of Alibaba Cloud’s Qwen 2.5 14B model, citing a high statistical correlation and architectural similarities supported by an unusual “fingerprinting” analysis. These allegations triggered intense debate over intellectual property rights, originality, and ethical practices in AI development, especially amid the competitive landscape of Chinese AI enterprises.
In response, Huawei’s Noah’s Ark Lab firmly denied the accusations, asserting that Pangu Pro MoE was independently developed using original architecture and trained exclusively on Huawei hardware. The lab acknowledged referencing some open-source code components in compliance with licensing norms but rejected claims of unauthorized use of Alibaba’s proprietary models. Experts have noted that observed similarities may arise from common design patterns within the AI community rather than direct copying, underscoring the complex technical and legal nuances in distinguishing AI model originality.
This controversy highlights broader challenges in the evolving legal and ethical frameworks surrounding AI-generated content and intellectual property. In China, recent judicial decisions and regulatory guidelines have begun addressing copyright and patent considerations for AI innovations, though uncertainties remain regarding authorship and liability. The Huawei-Alibaba dispute exemplifies the pressing need for clear policies that balance innovation incentives with respect for intellectual property rights in the rapidly developing AI sector.

Background

Huawei’s Pangu AI models represent a significant advancement in large language model (LLM) technology, developed entirely on Huawei’s Ascend AI chips. The Pangu Pro MoE (Mixture of Experts) 72B model, a hybrid expert architecture, exemplifies Huawei’s efforts to improve model efficiency and scalability by grouping experts to balance workload and optimize inference performance. This approach, known as Mixture of Grouped Experts (MoGE), addresses common inefficiencies observed in traditional MoE models where some experts are activated disproportionately, leading to system imbalance when distributed across devices.
Huawei’s AI lab, Noah Ark Lab, emphasized that the Pangu models were built from the ground up without incremental training on other manufacturers’ models, highlighting original architectural innovations and strict adherence to open-source licensing where third-party code was involved. The Pangu series, including PanGu-Σ and earlier iterations, were trained on large-scale clusters equipped with hundreds of Ascend 910 AI accelerators, processing vast multilingual datasets across natural and programming languages to ensure broad applicability and performance. The PanGu-Σ model notably incorporates Random Routed Experts (RRE) alongside Transformer decoder architectures, resulting in a training throughput significantly faster than comparable MoE models.
Huawei’s strategic release of the Pangu 7B dense model and Pangu Pro MoE 72B hybrid model under an open-source license marks a commitment to transparency and community engagement, providing developers with scalable options tailored to varying computational resources. These efforts are situated within China’s broader national AI development plan aimed at achieving global leadership in artificial intelligence by 2030, underlining the strategic importance of Huawei’s advancements in AI infrastructure and model innovation.
Furthermore, Huawei’s research on Pangu Pro MoE training techniques illustrates their success in overcoming core challenges related to massive MoE models on specialized hardware through systematic architecture search, efficient communication methods, and memory optimizations. This has allowed Pangu Pro MoE to outperform other prominent open-source models in the sub-100 billion parameter category, such as GLM-Z1-32B and Qwen3-32B. The models have been applied across diverse industries including finance, transportation, education, and logistics, showcasing their practical impact through Huawei Cloud’s intelligent services platform.

Allegations Regarding Source Code

In mid-2023, controversy arose concerning Huawei’s open-sourced Pangu Pro MoE large language model (LLM) after a group named HonestAGI published a report alleging that Pangu Pro MoE was not an original creation but rather an “upcycled” version of Alibaba’s Qwen 2.5 14B model. The report claimed an “extraordinary correlation” of 0.927 between Pangu and Qwen based on a novel fingerprinting technique, arguing that this statistical similarity was robust enough to persist despite further training, a common method used to remove traditional watermarks from stolen models. HonestAGI also highlighted architectural parallels such as nearly identical patterns in QKV bias projections and attention LayerNorm weights, and even discovered a Qwen license file within Pangu’s official code repository on GitCode.
These allegations sparked significant debate about the originality and intellectual property status of Huawei’s model. Multiple whistleblowers purportedly from within Huawei’s team corroborated claims that Pangu Pro MoE incorporated unauthorized elements from Alibaba’s Qwen model. This situation came at a time when Huawei was seeking to increase adoption of its AI technology by open-sourcing Pangu Pro MoE on GitCode in late June 2023, positioning itself competitively against other industry players like Alibaba and Meta.
In response, Huawei’s Noah’s Ark Lab clarified that the development of Pangu Pro MoE did reference certain open-source codes from other large language models, following industry open-source practices. The lab emphasized that they strictly complied with open-source license requirements and clearly labelled any incorporated code to respect intellectual property rights. This statement highlighted the common practice in the AI community of building upon existing open-source frameworks while adhering to legal and ethical standards.
The broader context of these allegations intersects with ongoing legal and regulatory discussions about copyright protection and authorship in AI-generated works. Under current laws, including those in China, copyright generally requires human intellectual creation and a modest degree of originality or creative effort, which AI-generated outputs alone may not fulfill. Court rulings, such as one by the Beijing Internet Court in 2023, have begun recognizing copyrightability for AI-generated images when a human prompt can be considered an authorial act, indicating evolving legal perspectives on AI authorship. This legal ambiguity underscores the challenges in addressing ownership and infringement claims relating to AI models like Pangu Pro MoE.

Huawei AI Lab’s Response

Huawei’s AI research division, Noah’s Ark Lab, responded firmly to allegations that its Pangu Pro MoE (Mixture of Experts) model had copied elements from Alibaba’s Qwen 2.5 14B model. The claims surfaced after an entity named HonestAGI published an analysis on GitHub highlighting an “extraordinary correlation” of 0.927 between the two models, along with architectural similarities such as near-identical QKV bias projections and attention LayerNorm weights. HonestAGI further pointed to the presence of a Qwen license file within Pangu’s official code repository, suggesting possible code reuse.
In a public statement issued via WeChat, Noah’s Ark Lab rejected these accusations, asserting that the Pangu Pro MoE model was independently developed and trained exclusively on Huawei’s homegrown Ascend hardware platform, including GPUs and NPUs. The lab emphasized that the model was not the product of incremental training on rival models and that their development process adhered strictly to open-source licensing requirements. They acknowledged that certain foundational components referenced open-source codes from other large language models (LLMs) following standard industry practices but insisted that these usages were clearly labeled and compliant with licensing terms.
Noah’s Ark Lab also highlighted the technical innovations of Pangu Pro MoE, describing it as the world’s first model of its kind trained on Ascend chips, which represent Huawei’s response to Nvidia’s AI accelerators. The model incorporates solutions for large-scale distributed training load balancing and enhanced training efficiency. Additionally, researchers developed the Pangu Ultra MoE, a sparse LLM with 718 billion parameters, optimized to leverage the capabilities of Ascend hardware through highly structured training methodologies.
While whistleblowers claiming affiliation with Huawei allegedly supported the plagiarism claims, the company maintained that the observed correlations were insufficient to prove copying, suggesting that similarities may result from common design choices within the AI community. Furthermore, some experts argued that diversity among AI models, as observed in comparisons between models like Qwen and other contemporaries, reflects innovation rather than imitation.

Technical Comparison and Distinctions

A detailed comparative analysis between Huawei’s Pangu Pro MoE model and other large language models such as Qwen 2.5-14B and Hunyuan A13B reveals significant architectural and representational differences. Visualizations show that hybrid-Hunyuan-A13B and Qwen 2.5-14B exhibit distinct internal patterns at various model levels, indicating different design philosophies and learned representations. This diversity highlights the innovation occurring across AI teams rather than mere imitation.
Huawei’s Pangu Pro MoE leverages the Mixture of Grouped Experts (MoGE) architecture, which improves upon traditional Mixture of Experts (MoE) by grouping experts during selection and balancing workload more efficiently. This design enhances computational load distribution across multiple devices and optimizes data transmission, particularly during inference. Such architectural choices enable Pangu Pro MoE 72B to outperform its predecessors in both speed and efficiency.
Further, the PanGu-Σ model integrates Random Routed Experts (RRE) within a Transformer decoder framework, facilitating the extraction of sub-models tailored for diverse applications like conversation, translation, and code generation. This architecture achieves a training throughput 6.3 times faster than comparable MoE models with the same hyperparameters. The Pangu team’s approach to large-scale model training involves systematic architecture search, efficient inter-device communication, and customized memory optimizations, which collectively support scalable AI training on Ascend GPUs and NPUs.
Although some foundational components of Pangu Pro MoE’s codebase reference open-source practices and incorporate portions of code from existing open-source LLMs, this is consistent with industry standards rather than plagiarism. This reuse of code is acknowledged as a common practice for foundational building blocks across AI projects.

Legal and Ethical Considerations

The development and deployment of generative AI models such as Huawei’s Pangu Pro MoE raise complex legal and ethical questions, particularly regarding intellectual property (IP) rights, copyright ownership, and data compliance. Courts and regulatory bodies are increasingly addressing these issues amid the rapid evolution of AI technologies.
In China, judicial decisions have begun to recognize that AI-generated works may qualify for copyright protection, provided that the claimant demonstrates a requisite level of intellectual creativity as mandated under Chinese law. The Supreme Court has emphasized that originality requires a modest degree of intellectual labor, going beyond mere data or factual compilation, thereby affirming the human-centric notion of authorship while acknowledging AI’s creative contributions. However, the question of whether AI itself can be regarded as an author remains unresolved, pending further technological and legal developments.
The China National Intellectual Property Administration’s “Guidelines for Patent Applications for AI (Trial Implementation)” released in December 2024 represent an effort to clarify legal standards for AI-related innovations and patent filings, underscoring the importance of respecting intellectual property rights and personal data protection during model training and use. Specifically, generative AI providers are required to source training data legally and ensure compliance with data privacy laws, including obtaining explicit consent where personal data is involved.
Secondary liability has also been established in cases where AI platform operators fail to implement reasonable preventive measures against copyright infringement by end-users. A recent court ruling held that defendants should anticipate infringing activities by their users and bear corresponding responsibilities, highlighting the necessity for AI developers and service providers to actively manage IP risks.
In the context of Huawei’s Pangu Pro MoE, allegations of unauthorized use of open-source code have surfaced, prompting scrutiny over compliance with open-source licenses and ethical development practices. Huawei’s Noah’s Ark Lab acknowledged incorporating certain open-source components but maintained that these were used in strict accordance with licensing requirements and properly attributed. These issues illustrate the broader challenges AI developers face in balancing innovation with respect for legal and ethical frameworks governing source code use and IP protection.
Judicial reforms in China, including mechanisms to ensure consistent rulings and the application of AI tools to promote uniformity in similar cases, further shape the evolving legal landscape impacting AI technologies. As the intersection of AI and IP law continues to develop, stakeholders must navigate a complex environment where both legal compliance and ethical considerations are paramount.

Impact and Reactions

The release of Huawei’s Pangu Pro MoE model has generated significant attention within the AI community and beyond, not only due to its technical advancements but also because of the legal and ethical controversies surrounding its development. The Pangu Pro MoE model, a 72-billion parameter hybrid expert system designed to optimize efficiency by activating only a subset of its parameters for each input, represents a breakthrough in large language model design, promising substantial increases in learning capacity with manageable execution costs.
However, shortly after its open-source release, Huawei faced allegations from rival Alibaba Cloud, which claimed that the Pangu Pro MoE was trained using Alibaba’s proprietary Qianwen Qwen-2.5 14B model, raising questions of copyright infringement and intellectual property violations. This accusation has sparked a wider debate on the legal status of AI-generated works and the applicability of existing copyright frameworks to AI models. Central to this debate is whether AI systems themselves can be recognized as authors or rights holders, and how liability should be attributed when AI-generated content potentially infringes on copyrighted material.
The legal uncertainty surrounding AI-generated works is further complicated by recent regulatory developments in China, such as the December 2024 “Guidelines for Patent Applications for AI (Trial Implementation)” issued by the China National Intellectual Property Administration. These guidelines aim to clarify some aspects of AI-related intellectual property, but statutory ambiguities remain, making judicial interpretations crucial for shaping the evolving landscape. Chinese courts, influenced by reforms enhancing judicial accountability and referencing of precedent, are expected to play a pivotal role in determining outcomes of such cases.
Industry and academic observers emphasize that the outcome of the Huawei-Alibaba dispute will have broad implications for AI development, innovation strategies, and intellectual property governance, especially in contexts where AI models are trained on datasets incorporating potentially copyrighted material. The controversy highlights the pressing need for clear policies and legal frameworks that balance the promotion of AI innovation with respect for existing intellectual property rights, ensuring responsible development and deployment of advanced AI technologies.

Timeline of Key Events

On June 30, Huawei officially made two of its Pangu AI models available to developers as part of an open-source strategy. These models included the Pangu 7B dense model and the Pangu Pro MoE 72B hybrid expert model, designed to provide scalable options for various computational needs and industry applications.
Shortly after the release, on July 6, controversy erupted when a group calling itself HonestAGI published a report on GitHub, which was later removed but remains accessible via the Web Archive. The report alleged that Huawei’s Pangu Pro MoE model was not an original creation but rather an “upcycled” version of Alibaba Cloud’s Qwen 2.5 14B model. This claim was based on a novel “fingerprinting” technique intended to demonstrate model derivation rather than independent training.
Following the allegations, Huawei’s AI lab strongly denied any accusations of plagiarism, rejecting the claims that its Pangu model was copied from Alibaba’s Qwen model. Huawei emphasized that their development process did not involve unauthorized use of competitor data or models.
The dispute prompted widespread discussion in AI research communities and Chinese technology media about the originality and development methodologies of large language models, particularly those employing Mixture of Experts (MoE) architectures that activate subsets of parameters during execution.
This episode has taken place amid a broader context of regulatory uncertainty concerning intellectual property ownership and AI-generated works. Court judgments and