Huawei denies claims its AI model copied Alibaba’s Qwen system

Huawei’s AI research unit has rejected allegations that its Pangu Pro Moe model copied Alibaba’s Qwen 2.5, insisting it was independently developed.

The statement by Huawei’s Noah Ark Lab came on Saturday, a day after an entity named HonestAGI published a paper on GitHub claiming “extraordinary correlation” between Huawei’s Pangu Pro Moe and Alibaba’s Qwen 2.5-14B model.

The paper alleged that Huawei’s model was derived through “upcycling” rather than being trained from scratch, and suggested potential copyright violations and misleading claims about Huawei’s development investment.

Noah Ark Lab said its Pangu model was “not based on incremental training of other manufacturers’ models” and had introduced “key innovations in architecture design and technical features.”

It added that Pangu Pro Moe was the first large-scale model fully trained on Huawei’s Ascend chipsets.

The lab also said its team had complied with open-source licence terms for any third-party code used, though it did not specify which models had informed its development.

Alibaba, which released the Qwen 2.5-14B model in May 2024, did not respond to a Reuters request for comment. Reuters was unable to identify or contact HonestAGI.

Huawei first entered the large language model market in 2021 but has since been seen as trailing behind rivals. Its Pangu Pro Moe model was open-sourced via China’s GitCode platform in late June in an effort to attract external developers.

While Alibaba’s Qwen series targets consumer applications, such as chatbots, Huawei’s Pangu models are typically geared towards enterprise and government use, particularly in finance and manufacturing.

Copied