Huaweis AI lab denies that one of its Pangu models copied Alibabas Qwen

In a strong rebuttal to recent claims, Huawei's artificial intelligence research division has defended its Pangu Pro large language model, asserting that it was developed independently and not derived from Alibaba's model. The statement from Noah Ark Lab came in response to a paper released by an organization named HonestAGI, which alleged that Huawei's Pangu Pro Moe model exhibited significant similarities to Alibaba's Qwen 2.5 14B. According to HonestAGI, the findings indicated that Huawei's model may have been created through a process they termed "upcycling," suggesting it was not built from the ground up. This sparked considerable debate within AI communities online and across Chinese tech media. The paper also implied potential copyright infringements and accused Huawei of fabricating details regarding its investment in the model's training. In its defense, Noah Ark Lab stated emphatically that their model was not based on incremental improvements of other companies' models. They highlighted the innovations in both architecture and technical features that set Pangu Pro apart. Additionally, they noted that this model is the first large-scale one developed entirely on Huawei's proprietary Ascend chips. The lab further asserted compliance with open-source licensing for any third-party code utilized, though specifics on which open-source models influenced their design were not disclosed. Alibaba has yet to respond to inquiries regarding these claims. The situation has unfolded against the backdrop of increasing competition in the AI sector, particularly highlighted by the launch of DeepSeek's open-source model R1 earlier this year, which captured attention due to its affordability. Meanwhile, Alibaba's Qwen 2.5-14B, released in May 2024, is part of a model family aimed at consumer applications, contrasting Huawei's focus on sectors such as government and finance with its Pangu models. While Huawei has been perceived as trailing behind its competitors since its initial foray into large language models with Pangu in 2021, it has sought to regain ground by open-sourcing its Pangu Pro Moe models on GitCode to enhance developer engagement and adoption of its AI technologies.