On July 5th, Huawei Noah's Ark Lab released a statement stating that the Pangu Pro MoE open-source model is a basic large model developed and trained based on the Ascend hardware platform, and is not based on incremental training of models from other vendors.
It is reported that on June 30th, Huawei announced the open source of Pangu 7B dense and 72B hybrid expert models, which were suspected of plagiarizing the Alibaba Cloud Tongyi Qianwen Qwen-2.5 14B model. The author who questioned the plagiarism of Huawei's Pangu team pointed out in the paper that Pangu's MoE model and Qwen-2.514B model have a very high correlation in the distribution of attention parameters (with an average correlation of 0.927), far exceeding the normal range compared to other models.
In response, the Pangu team stated in a statement that the code implementation of some basic components of the Pangu Pro MoE open source model referenced industry open source practices and involved some open source code of other open source large models. The Pangu team strictly follows the requirements of the open source license and clearly marks the copyright statement of the open source code in the open source code files. This is not only a common practice in the open source community, but also in line with the open source collaboration spirit advocated by the industry.
