Huawei’s AI division, Noah’s Ark Lab, has firmly denied allegations that its Pangu Pro Moe large language model (LLM) copied elements from Alibaba’s Qwen 2.5 14B model. Huawei has directly refuted a publication released by an obscure group HonestAGI claiming that the newly launched model of the firm follows strict intellectual property procedures. The HonestAGI paper asserts that Huawei’s product, possibly upcycled from Alibaba’s Qwen 2.5-14B, sparked questions about copyright violation, case transparency, and originality amid China’s intensifying AI race.

It has rapidly become one of the pivotal points in the larger ongoing conversation of intellectual rights with regards to the development of proportionate AI models, specifically in China where the level of competition is especially high in its technology-driven sector. The authenticity of model-trained practices has been under extra consideration as main players attempt to bring forth forms that are competent, effective and expandable in consumer and business opportunities.

Background: The Rise of Pangu and the AI Rivalry in China

Huawei entered the large language model scene with its 2021 launch of Pangu, aiming to establish itself as a key player in enterprise-level artificial intelligence. Since this debut, Chinese competitors Alibaba, Baidu, and Tencent have accelerated their development pace, often outshining Huawei’s progress.

Meanwhile, Alibaba’s Qwen series, particularly the Qwen 2.5 family released in May 2024, has emerged as a significant low-powered yet high-performance architecture, tailored specifically for consumer devices like PCs and smartphones. Hyped as highly able to customize the conversations, being technologically efficient, the Qwen models are similar to the chatbot- styled interactions led by OpenAI and ChatGPT in both aspects of form, as well as functionality.

Comparatively, Huawei had moved towards industrial and governmental usage with a focus on its application in finance, manufacturing and digital-city infrastructure, promoting its Pangu. Under this focus, however, critics argue that Huawei has lagged behind in terms of public recognition as well as innovativeness especially with the advent of open-source rivals like DeepSeek R1 model attracted the interest of the global community with its exemplary performance at relatively cheap prices.

The Controversy: HonestAGI’s Allegation and its Implications

On July 5, 2025, HonestAGI released a research paper on GitHub that accused Huawei of not training Pangu Pro Moe from scratch. The paper reports that Huawei’s model exhibits “extraordinary correlation” with Alibaba’s Qwen 2.5-14B, prompting suspicions that Huawei retrained or fine-tuned Alibaba’s model to create an “upcycled” version, rather than developing it independently.

Key claims from HonestAGI’s report include:

• Identical parameter initialization patterns in certain layers of the model.
  
• Highly similar token output probabilities under identical prompts.
  
• Lack of deviation in architectural choices that would suggest independent research.
  
• Potential fabrication of training reports, suggesting misrepresentation of Huawei’s R&D efforts.
  

These assertions, if proven, could imply a violation of copyright, misuse of open-source licenses, and reputational damage for Huawei, particularly in international markets where transparency and compliance are non-negotiable.

Huawei Responds: Full Denial, Focus on Independence and Innovation

Huawei’s Noah Ark Lab issued a rebuttal shortly after the allegations surfaced. In its official statement, the lab insisted that Pangu Pro Moe was developed independently, with “key innovations in architecture design and technical features.”

The statement emphasized:

• The model was not trained on or derived from any other manufacturer’s model.
  
• It was built entirely on Huawei’s own Ascend AI chips, showcasing its proprietary AI hardware-software integration.
  
• Any open-source code used was done so in compliance with license requirements, though the lab did not name specific models or repositories referenced.
  

Noah Ark Lab also highlighted the strategic shift Huawei has taken with the recent open-sourcing of the Pangu Pro Moe model on GitCode—China’s answer to GitHub—in late June 2025. Huawei presented this move as part of a broader initiative to enhance adoption, promote transparency, and engage the developer community in shaping Pangu’s future.

Alibaba’s Silence and HonestAGI’s Anonymity

Interestingly, Alibaba has not issued a public statement on the matter, and Reuters reported being unable to contact HonestAGI, the group or individual behind the original accusation. This lack of visibility raises questions about the credibility and motivation of the whistleblower, although the technical rigor of the report has led many AI researchers to take it seriously.

The silence from Alibaba could reflect internal investigations, strategic restraint, or legal caution. However, industry insiders speculate that Alibaba may opt to pursue quiet resolution or leverage the controversy to strengthen its market position, especially given its higher consumer visibility through chatbot applications.

AI Race and Intellectual Property Challenges in China

The Huawei-Alibaba conflict highlights deeper tensions within the AI ecosystem, including how to establish model novelty and enforce intellectual property (IP) into a world where matters of open-source weights, overlapping architectures, and converging design philosophies all contribute to fail to distinguish one AI programme or methodology against another.

As China continues to struggle to gain international competitiveness within the field of AI, local firms are coming under increasing pressure to roll-out a new model of excellence as quickly as possible. In this context, accusations of shortcutting development through model cloning or fine-tuning existing models without proper attribution are becoming more frequent.

This isn’t the first time the LLM community has faced such scrutiny. In late 2023, several U.S.-based AI startups were accused of using GPT-3 derivatives to claim original development. OpenAI, Meta, and other leading labs have since called for clearer guidelines and standardized auditing mechanisms to trace model training origins.

What’s Next for Huawei and the Industry

The fallout from this controversy may be a turning point for model governance in China, particularly as local models aim to scale internationally. Huawei may need to:

• Release detailed training logs or datasets to validate its claims.
  
• Allow independent audits of the Pangu Pro Moe model lineage.
  
• Engage more actively with open-source communities to foster transparency and trust.
  

Meanwhile, policymakers in China and abroad may consider developing a framework akin to a “Model Provenance Certification”—a digital trail that records a model’s training process, datasets, and architectural lineage, similar to a software supply chain bill of materials (SBOM).

Conclusion: A Battle of Trust and Technological Sovereignty

As large language models continue to redefine the digital landscape, the question of who built what is no longer just academic. It touches on national AI strategy, commercial integrity, developer trust, and ethical deployment of increasingly powerful models.

The latest announcement of Huawei has, in the meantime, brought an end to the rumours about its collaboration with Alibaba, although until confirmation by external sources, or even a formal comment by Alibaba itself, the technology world will carry on proceeding with a tint of uncertainty. Both developers and enterprises, therefore, should challenge AI models not only in terms of performance but also of provenance, transparency and compliance.

In a global market that rewards innovation but punishes opacity, trust may soon be the most valuable parameter of all.