Summary
Unlocking the Truth: Huawei’s AI Lab Responds to Allegations about Pangu Pro MoE Model Source Code details the controversy surrounding Huawei’s large language model (LLM), Pangu Pro MoE, developed by its Noah’s Ark Lab. The Pangu Pro MoE is a 72-billion-parameter Mixture of Experts (MoE) model leveraging Huawei’s proprietary Ascend AI hardware and an innovative Mixture of Grouped Experts (MoGE) architecture, designed to improve computational efficiency and scalability in sparse expert models. It has been positioned by Huawei as a leading indigenous AI innovation that outperforms several prominent open-source counterparts under 100 billion parameters, marking a significant milestone in China’s AI development efforts.
In mid-2025, the model became the subject of serious allegations from the open-source community, particularly by the GitHub account HonestAGI, which claimed that Pangu Pro MoE’s source code bore “extraordinary correlation” with Alibaba’s Qwen-2.5 14B model, suggesting possible plagiarism or unauthorized reuse of proprietary elements. The analysis pointed to striking similarities in technical components such as QKV bias projections and attention layer normalization weights, with whistleblowers reportedly from within Huawei corroborating these concerns. This sparked widespread debate over intellectual property rights, ethical research practices, and the challenges of distinguishing legitimate open-source collaboration from potential IP infringement in the rapidly evolving AI landscape.
Huawei’s Noah’s Ark Lab responded by acknowledging the use of “certain open-source codes” compliant with licensing requirements, while firmly denying that the Pangu Pro MoE was derived from incremental training on rival models. The company emphasized the originality of its MoGE architecture and the exclusive use of Huawei’s Ascend hardware platform in training the model. Nonetheless, internal dissent within the lab and continued scrutiny from the AI research community have underscored the complexities and sensitivities around transparency, accountability, and innovation integrity in China’s AI sector.
The controversy surrounding Pangu Pro MoE not only highlights technical and ethical issues in large language model development but also resonates within broader geopolitical and legal contexts involving technology transfer, intellectual property, and national security concerns. It exemplifies the challenges faced by leading technology companies in balancing rapid AI innovation with ethical standards and intellectual property rights amid increasing international competition and scrutiny.
Background
Huawei’s Pangu Pro MoE (Mixture of Experts) model represents a significant advancement in large language model (LLM) development, particularly in efficiently scaling sparse expert architectures. Developed by Huawei’s AI research division, Noah Ark Lab, the Pangu Pro MoE is a 72-billion-parameter hybrid expert model that utilizes a novel Mixture of Grouped Experts (MoGE) architecture. This design groups experts during selection and ensures balanced activation across expert groups, which mitigates computational load imbalance across devices by over 50% and enhances throughput during both training and inference phases. Approximately 16 billion parameters are activated per token, while the model’s sparse expert modules contain 95% of total parameters, with attention mechanisms accounting for the remaining 5%.
The model was trained exclusively on Huawei’s Ascend GPUs and NPUs, leveraging specialized hardware optimizations and adaptive pipeline overlap techniques to maximize communication-computation overlap and system efficiency. These innovations contribute to an excellent cost-to-performance ratio, with inference speeds reaching up to 1528 tokens per second per Ascend 8001 A2 card. The Pangu Pro MoE is considered a forefront model under the 100 billion parameter scale, outperforming prominent open-source competitors such as GLM-Z1-32B, Qwen3-32B, and Gemma3-27B across multiple benchmarks.
Huawei has emphasized that the Pangu Pro MoE is the first large-scale language model built entirely on its Ascend chips and that the development process involved key architectural and technical innovations independently devised by Noah Ark Lab. This assertion was made in response to public allegations suggesting that elements of the Pangu Pro MoE were derived from rival models, notably Alibaba’s Qwen series. Huawei categorically denied these claims, highlighting that the model was not based on incremental training of other manufacturers’ models and underscoring its proprietary architectural advancements.
The controversy surrounding the Pangu Pro MoE has drawn attention to broader issues of intellectual property and innovation integrity within China’s rapidly growing AI sector. It has also sparked a dialogue about the challenges of distinguishing legitimate open-source collaboration from potential IP infringements in the competitive landscape of large language model development.
Allegations Regarding the Source Code
In mid-2025, Huawei’s Noah’s Ark Lab faced significant controversy concerning the source code of its Pangu Pro MoE (Mixture of Experts) large language model. An entity known as HonestAGI published an analysis on the code-sharing platform GitHub, alleging that Pangu Pro MoE exhibited an “extraordinary correlation” with Alibaba’s Qwen 2.5 14B model. Specifically, HonestAGI highlighted striking similarities in the QKV bias projections and attention layer normalized weights between the two models, revealing a correlation coefficient as high as 0.927, which suggested potential plagiarism rather than coincidental design overlap.
HonestAGI’s claims were bolstered by multiple whistleblowers purportedly from Huawei’s internal team, who confirmed the allegations against Pangu Pro MoE and criticized the lab’s internal response, with at least one insider resigning over perceived complicity in misleading reports. Despite these accusations, Noah’s Ark Lab responded by acknowledging that the development of Pangu Pro MoE incorporated “certain open-source codes” from other open-source large language models, emphasizing strict adherence to open-source licensing requirements and clear attribution of the sourced code. The lab maintained that their use of open-source components was legitimate and transparent, distinguishing their practices from improper intellectual property usage.
This dispute has drawn considerable attention within the AI research community, raising broader questions about transparency, licensing compliance, and ethical collaboration in the rapidly evolving landscape of Chinese large language model development. The allegations and ensuing responses underscore the challenges of distinguishing between legitimate open-source collaboration and intellectual property violations in AI model development.
Huawei AI Lab’s Response
In response to allegations raised by the open-source community HonestAGI, which claimed that Huawei’s Pangu Pro MoE 72B model exhibited “extraordinary correlation” with Alibaba’s Qwen-2.5 14B model and suggested possible copyright violations, Huawei’s Noah’s Ark Lab issued a detailed statement defending the integrity of their work. The Lab acknowledged that the development of the Pangu Pro MoE model involved the use of “certain open-source codes” from other models but emphasized strict compliance with open-source license requirements and transparent labelling of such code.
Huawei clarified that the Pangu Pro MoE model was developed and trained on their proprietary Ascend AI hardware platform and was not the result of incremental training or direct derivation from any existing models, including Qwen-2.5. They highlighted the innovative MoGE (Mixture of Grouped Experts) architecture they introduced, which improves expert workload balancing and overall training efficiency, underscoring the technical advancements embedded in Pangu Pro MoE beyond prior versions. Furthermore, Huawei’s team pointed out their systematic approach to tackling the challenges of training large-scale Mixture-of-Experts models on specialized hardware, reinforcing the originality and scalability of their architecture.
Despite Huawei’s rebuttal, the controversy attracted considerable attention within AI and Chinese tech communities due to the nuanced distinction between legitimate open-source collaboration and potential intellectual property concerns in AI model development. Huawei’s lab maintained transparency in their communication, including responding to technical clarifications on intranet platforms to address community questions directly. The company reiterated its commitment to openness by making the Pangu AI models available to developers, aiming to foster trust and collaboration in the AI research ecosystem.
Nevertheless, internal dissent surfaced as some individuals reportedly expressed concerns about the handling of the controversy and alleged falsehoods in official reports related to the model’s development. This underscored the complexities Huawei faced in navigating both external scrutiny and internal accountability during the dispute.
Technical Analysis of the Models
The Pangu Pro MoE model employs an innovative architecture known as the Mixture of Grouped Experts (MoGE), which enhances the traditional Mixture of Experts (MoE) framework by grouping experts during selection to balance the workload more effectively. This design ensures a balanced computational load when model execution is distributed across multiple devices, significantly improving data transmission and throughput, especially during the inference phase. The model activates 16 billion parameters per token out of a total of 72 billion, achieving a superior cost-to-performance ratio on the Ascend 300I Duo hardware platform. This architectural choice enables Pangu Pro MoE to outperform other prominent open-source models under 100 billion parameters, such as GLM-Z1-32B and Qwen3-32B, across a variety of competitive benchmarks.
However, the technical analysis has also revealed contentious points regarding the originality of Pangu Pro MoE’s design. Detailed investigations comparing architectural patterns, specifically focusing on QKV bias projections and attention layer normalization weights, have exposed striking similarities between Pangu Pro MoE and the Qwen2.5-14B model. This similarity extends across all three projection types (query, key, and value), raising concerns that these resemblances surpass coincidental design choices and suggest a deeper connection between the models. Additional analysis from independent researchers, including HonestAGI, has supported these findings, pointing to overlapping technical characteristics that have fueled allegations of plagiarism against Pangu Pro MoE. Multiple whistleblowers purportedly from Huawei’s team have corroborated these concerns, further intensifying the scrutiny on the model’s source code and architecture.
Despite these controversies, Huawei’s Pangu Pro MoE remains a leading example of large-scale sparse model training on Ascend NPUs, leveraging massive parallelization and sparsity to push performance boundaries in sub-100B parameter LLMs. The Mixture of Experts architecture continues to gain traction in the field due to its ability to scale model capacity efficiently while managing computational costs. Yet, the dispute over design originality underscores ongoing challenges in transparency and intellectual property within the rapidly evolving landscape of large language models.
Impact and Reactions
The revelations concerning the Pangu Pro MoE model’s source code have generated significant controversy within the AI research community and beyond. The core of the dispute lies in allegations that the Pangu Pro MoE incorporated open-source components from other large language models without proper attribution, raising concerns about intellectual property rights and ethical standards in AI development. This has intensified scrutiny on the integrity of China’s large language model development efforts, highlighting the fine line between collaborative open-source innovation and potential code misappropriation.
Within the developer community, the controversy has sparked considerable debate and led to investigations into the architectural similarities between Pangu and other models. Members of the open-source LLM community have actively engaged in examining architectural patterns such as QKV bias projections and attention layer normalization weights, revealing notable overlaps with models like Pangea and Qwen2.5-14B. These findings have fueled ongoing discussions about transparency and originality in AI model design.
The controversy also triggered internal turmoil within the team behind Pangu Pro MoE. Reports indicate that at least one key contributor publicly resigned, citing ethical concerns and the pressure to approve misleading reports about the model’s development. This resignation has been viewed as a sign of internal dissent and a call for greater accountability in AI research practices.
Beyond the AI community, the incident has geopolitical implications, contributing to the growing tensions around technology and national security. Historical precedents, such as the U.S. invoking Cold War-era powers to investigate Chinese cyber activities in telecommunications, underscore the sensitivity surrounding Chinese technological advancements and foreign scrutiny. Consequently, the Pangu Pro MoE controversy is situated within a broader context of international distrust and competition in emerging technologies.
Despite these challenges, proponents argue that the Pangu Pro MoE architecture offers significant efficiency advantages, such as an improved cost-to-performance ratio in inference tasks using the Ascend 3001 Duo hardware. This technical merit emphasizes the importance of balancing innovation benefits with ethical development and transparency in the AI field.
Legal and Ethical Considerations
The release of Huawei’s Pangu Pro MoE model sparked significant legal and ethical debates within the AI community and the broader technology industry. Central to these concerns were allegations of intellectual property violations, with claims that Pangu Pro MoE bore extraordinary similarities to Alibaba’s Qwen-2.5 14B model. These accusations originated from a research paper published on GitHub by the account HonestAGI, which suggested that Huawei’s model might have involved unauthorized use or “upcycling” of Alibaba’s proprietary technology rather than being independently developed from scratch.
In response, Huawei’s Noah’s Ark Lab emphasized that Pangu Pro MoE was independently developed and trained on Huawei’s homegrown Ascend AI chips and Ascend hardware platform. They acknowledged the incorporation of certain open-source code segments from other large language models but asserted that all such use strictly complied with open-source licensing requirements, with appropriate code labeling and attribution. The lab also refuted claims of incremental training based on other models, maintaining that their work represented an original contribution to AI development.
This controversy reflects broader ethical questions about the boundaries between legitimate open-source collaboration and potential intellectual property infringement in AI development. The distinction becomes particularly critical given the rapid evolution of large language models and the prevalent practice of integrating open-source components to accelerate innovation. Moreover, the debate highlights the challenges of transparency and accountability in AI research, especially when high-profile corporations are involved.
The legal landscape surrounding these issues is further complicated by national security considerations, exemplified by prior regulatory actions such as the 2018 Federal Communications Commission (FCC) decision to restrict government subsidies for telecom equipment from companies perceived as national security risks, including Huawei. While this context is not directly related to the Pangu Pro MoE dispute, it underscores the geopolitical sensitivities intertwined with technological advancements and intellectual property rights involving Huawei.
Technically, Huawei’s Pangu Pro MoE model utilizes a Mixture of Grouped Experts (MoGE) architecture designed to optimize workload distribution and inference efficiency across Ascend NPUs. This approach leverages sparsity and expert grouping to balance computational load, representing a sophisticated advancement in large-scale AI model design. Despite this, the legal and ethical scrutiny primarily centers on the provenance and originality of the underlying source code and training methodologies.
Future Developments
Looking ahead, Huawei’s AI lab is expected to intensify efforts to clarify the development origins and technical foundations of the Pangu Pro MoE model to restore confidence in its proprietary innovations. Despite the current controversy surrounding allegations of source code borrowing and potential intellectual property violations, Huawei has emphasized its commitment to advancing indigenous AI technologies, particularly leveraging its Ascend AI hardware platform for training and deployment.
To address these concerns, future initiatives may include increased transparency in model training processes and more rigorous documentation to distinguish legitimate open-source collaboration from unauthorized usage. There is also potential for Huawei to further optimize the Pangu Pro MoE’s performance, given its current architecture as a sparse mixture-of-experts model with 72 billion parameters and strong benchmark results against competitive open-source models.
Moreover, ongoing discussions within the AI community and regulatory scrutiny could shape Huawei’s approach to compliance and intellectual property management, potentially influencing its research and development strategy. These developments will likely impact both Huawei’s positioning in the global AI landscape and the broader evolution of large language model innovation in China.
The content is provided by Jordan Fields, 11 Minute Read
