Huawei Denies Copying Alibaba’s Qwen AI in Pangu Models

Huawei’s AI research division, Noah Ark Lab, has vehemently denied recent allegations that its Pangu Pro large language model incorporates copied elements from Alibaba’s Qwen 2.5 14B model. The strong denial comes in response to a paper posted by an entity named HonestAGI on GitHub, which claimed “extraordinary correlation” between the two models, suggesting “upcycling” rather than independent development by Huawei.

The HonestAGI paper, which quickly gained traction in AI circles and Chinese tech media, further alleged potential copyright violation, misreporting of training efforts, and false claims regarding Huawei’s investment in the Pangu Pro Moe (Mixture of Experts) model.

In a statement issued over the weekend, Noah Ark Lab asserted that its Pangu Pro model was “independently developed and trained,” emphasizing “key innovations in architecture design and technical features.” The company also highlighted that Pangu Pro Moe is the first large-scale model built entirely on Huawei’s proprietary Ascend chips, a significant point of differentiation in the highly competitive AI landscape. Furthermore, Huawei stated that its development team strictly adhered to open-source license requirements for any third-party code utilized, though it did not specify which open-source models were referenced.

The accusations underscore the intensifying competition within China’s burgeoning AI sector, where tech giants are aggressively developing and deploying advanced models. Alibaba’s Qwen 2.5-14B, launched in May 2024, is optimized for consumer-facing applications like PCs and smartphones. In contrast, Huawei’s Pangu models, first introduced in 2021 and recently updated with Pangu Model 5.0 in June 2024, have primarily focused on enterprise-grade solutions for sectors such as government, finance, and manufacturing. Huawei open-sourced its Pangu Pro Moe models in late June on GitCode, aiming to boost adoption of its AI technology.

This dispute also brings to the forefront the evolving landscape of AI intellectual property protection in China. Recent landmark rulings in Chinese courts have begun to define copyright protection for AI-generated content and the boundaries of legitimate model development versus unauthorized copying. As the legal framework continues to mature, companies will likely face increasing scrutiny regarding the transparency of their AI training methodologies and the originality of their models.