Noah Ark Lab said in its statement that the model was 'not based on incremental training of other manufacturers' models' and that it had 'made key innovations in architecture design and technical features'. It is the first large-scale model built entirely on Huawei's Ascend chips, it added. — Reuters
BEIJING/SHANGHAI: Huawei's artificial intelligence research division has rejected claims that a version of its Pangu Pro large language model has copied elements from an Alibaba model, saying that it was independently developed and trained.
The division, called Noah Ark Lab, issued the statement on July 5, a day after an entity called HonestAGI posted an English-language paper on code-sharing platform Github, saying Huawei's Pangu Pro Moe (Mixture of Experts) model showed "extraordinary correlation" with Alibaba's Qwen 2.5 14B.
This suggests that Huawei's model was derived through "upcycling" and was not trained from scratch, the paper said, prompting widespread discussion in AI circles online and in Chinese tech-focused media.
The paper added that its findings indicated potential copyright violation, the fabrication of information in technical reports and false claims about Huawei's investment in training the model.
Noah Ark Lab said in its statement that the model was "not based on incremental training of other manufacturers' models" and that it had "made key innovations in architecture design and technical features." It is the first large-scale model built entirely on Huawei's Ascend chips, it added.
It also said that its development team had strictly adhered to open-source license requirements for any third-party code used, without elaborating which open-source models it took reference from.
Alibaba did not immediately respond to a Reuters request for comment. Reuters was unable to contact HonestAGI or learn who is behind the entity.
The release of Chinese startup DeepSeek's open-source model R1 in January this year shocked Silicon Valley with its low cost and sparked intense competition between China's tech giants to offer competitive products.
Qwen 2.5-14B was released in May 2024 and is one of Alibaba's small-sized Qwen 2.5 model family which can be deployed on PC and smartphones.
While Huawei entered the large language model arena early with its original Pangu release in 2021, it has since been perceived as lagging behind rivals. It open-sourced its Pangu Pro Moe models on Chinese developer platform GitCode in late June, seeking to boost the adoption of its AI tech by providing free access to developers.
While Qwen is more consumer-facing and has chatbot services like ChatGPT, Huawei's Pangu models tend to be more used in government as well as the finance and manufacturing sectors. – Reuters
