AI Lab Of Huawei Refutes That One Of Its Pangu Models Copied Alibaba’s Qwen

AI Lab Of Huawei Refutes That One Of Its Pangu Models Copied Alibaba’s Qwen

According to Noah Ark Lab, Pangu Proe Moe is the first large-scale model built entirely on Huawei’s Ascend chips.

The artificial intelligence (AI) research division of Huawei has refused claims that a version of its Pangu Pro large language model has imitated elements from an Alibaba model, saying that the said version was independently developed and trained. A statement was issued by the AI division of Huawei, Noah Ark Lab, a day after ‘Honest AGI’ posted a paper in the English language on code-sharing platform Github saying Pangu Proe Moe of Huawei showed ‘extraordinary correlation’ with Alibaba’s Qwen. The paper said that Huawei’s model was obtained through ‘upcycling’ and wasn’t trained from scratch. Its findings indicated potential copyright violation, fabrication of information in technical reports, and false claims about Huawei’s investment in training the model, the paper said.

Whereas Noah Ark Lab in its statement said that the model ‘is not based on incremental training of other manufacturers’ models’ and that ‘it has made key innovations in architecture design and technical features’. As per the statement, this is the first large-scale model built entirely on Huawei’s Ascend chips. The development team had strictly adhered to open-source licence requirements for any third-party code used, as per the statement.