BEIJING/SHANGHAI (Reuters) -Huawei’s synthetic intelligence analysis division has rejected claims {that a} model of its Pangu Professional giant language mannequin has copied components from an Alibaba mannequin, saying that it was independently developed and educated.
The division, known as Noah Ark Lab, issued the assertion on Saturday, a day after an entity known as HonestAGI posted an English-language paper on code-sharing platform Github, saying Huawei’s Pangu Professional Moe (Combination of Specialists) mannequin confirmed “extraordinary correlation” with Alibaba’s Qwen 2.5 14B.
This means that Huawei’s mannequin was derived by way of “upcycling” and was not educated from scratch, the paper mentioned, prompting widespread dialogue in AI circles on-line and in Chinese language tech-focused media.
The paper added that its findings indicated potential copyright violation, the fabrication of knowledge in technical studies and false claims about Huawei’s funding in coaching the mannequin.
Noah Ark Lab mentioned in its assertion that the mannequin was “not based mostly on incremental coaching of different producers’ fashions” and that it had “made key improvements in structure design and technical options.” It’s the first large-scale mannequin constructed completely on Huawei’s Ascend chips, it added.
It additionally mentioned that its improvement group had strictly adhered to open-source license necessities for any third-party code used, with out elaborating which open-source fashions it took reference from.
Alibaba didn’t instantly reply to a Reuters request for remark. Reuters was unable to contact HonestAGI or study who’s behind the entity.
The discharge of Chinese language startup DeepSeek’s open-source mannequin R1 in January this 12 months shocked Silicon Valley with its low price and sparked intense competitors between China’s tech giants to supply aggressive merchandise.
Qwen 2.5-14B was launched in Might 2024 and is one in all Alibaba’s small-sized Qwen 2.5 mannequin household which could be deployed on PC and smartphones.
Whereas Huawei entered the massive language mannequin enviornment early with its authentic Pangu launch in 2021, it has since been perceived as lagging behind rivals. It open-sourced its Pangu Professional Moe fashions on Chinese language developer platform GitCode in late June, in search of to spice up the adoption of its AI tech by offering free entry to builders.
Whereas Qwen is extra consumer-facing and has chatbot providers like ChatGPT, Huawei’s Pangu fashions are typically extra utilized in authorities in addition to the finance and manufacturing sectors.
(Reporting by Che Pan,Liam Mo and Brenda Goh; Enhancing by Edwina Gibbs)