Huawei’s AI research division, Noah Ark Lab, has firmly denied allegations that its Pangu Pro Moe model copied elements from Alibaba’s Qwen 2.5 14B model. The lab released a statement on Saturday insisting that its large language model was independently developed and trained, rejecting claims of “upcycling” raised in a report by an entity named HonestAGI, which alleged strong similarities between the two models.
The controversy began after HonestAGI published a paper on GitHub suggesting that Huawei’s Pangu Pro Moe showed “extraordinary correlation” with Alibaba’s model, implying it may not have been trained from scratch. The paper also accused Huawei of potential copyright violations, fabricating technical documentation, and exaggerating its resource investment. These claims have since sparked heated discussions among AI enthusiasts and in Chinese tech media.
In response, Noah Ark Lab emphasized that Pangu Pro Moe was the first large-scale model built entirely on Huawei’s Ascend chips, featuring original architecture and technical innovations. The lab also claimed it complied strictly with open-source licensing requirements, though it did not clarify which open-source materials, if any, were referenced during development. Alibaba has not publicly responded to the allegations, and the identity behind HonestAGI remains unknown.
Huawei, which has been seen as trailing behind competitors in the AI race, open-sourced its Pangu Pro Moe models on GitCode in late June 2025 to promote wider adoption. While Alibaba’s Qwen series targets consumer use and chatbot applications, Huawei’s Pangu line is primarily applied in sectors like government, finance, and manufacturing. The rivalry underscores the growing intensity of AI competition in China, especially following the emergence of cost-effective models like DeepSeek’s R1 earlier this year.