Huawei Denies AI Model Plagiarism Allegations Amidst Competitive Market Pressure

Huawei has publicly denied allegations that its open-source model Pangu Pro MoE is a «repurposed product» derived from Alibaba’s developments.

In an official statement from Huawei’s Noah’s Ark Lab, the company asserts that the Pangu Pro MoE model was entirely developed in-house and trained from scratch using Huawei Ascend hardware. They emphasize that the model was not created through fine-tuning on another vendor’s model.

These statements were made following an analysis by HonestAGI, which was removed from GitHub but remains accessible via the Wayback Machine. The report, titled «The Internal Fingerprint of Large Language Models: Fine-Tuning is NOT All You Need for Model Theft!» highlights significant similarities between Huawei’s Pangu Pro MoE and Alibaba’s Tongyi Qianwen Qwen-2.5 14B, particularly in the attention parameter distribution.

An anonymous analysis from GitHub suggested that Huawei’s model may have been «repurposed» rather than built from the ground up. Huawei has rejected this claim, stating that while some code employs «industry-standard open-source practices,» all open-source components are properly licensed and attributed.

These allegations come at a challenging time for Huawei as the company seeks to demonstrate that its Pangu models, which are trained on Ascend chips, can compete with global leaders in artificial intelligence. This is especially crucial as China looks for alternatives to Nvidia in light of tightening U.S. export restrictions.

For China, Huawei’s chips represent a pathway to navigate the U.S. government’s restrictions on Nvidia GPUs. The consequences of these limitations are already being felt: reports indicate that the release of the Deepseek R2 model is delayed due to a lack of access to Nvidia H20 chips.

Huawei is promoting its upcoming Ascend 910C processor as the leading alternative to Nvidia in China. Chip analyst Dylan Patel notes that the performance gap between the 910C and the previously approved Nvidia H20 has narrowed to less than a year compared to the two-year gap for the earlier 910B. China still lacks alternatives to Nvidia’s Blackwell chips, though reports suggest a scaled-down version for China is in development.

*Want to stay updated on significant AI news? Subscribe to our Telegram channel* [*BotHub AI News*](https://t.me/bothub).

For more details, you can find the original source of the news [here](https://the-decoder.com/huawei-pushes-back-on-ai-model-plagiarism-claims/).