TLDR: Huawei’s AI research division, Noah’s Ark Lab, has vehemently denied accusations from Alibaba Cloud that its newly open-sourced Pangu Pro MoE 72B large language model copied Alibaba’s Qwen-2.5. Huawei asserts its model was developed independently on its Ascend hardware, adhering strictly to open-source practices.
Huawei’s AI research arm, the Noah’s Ark Lab, has issued a strong denial against allegations from rival Alibaba Cloud, which claimed that Huawei’s recently open-sourced Pangu Pro MoE 72B hybrid expert model was trained using Alibaba’s proprietary Qwen-2.5 14B model. The controversy, which reignited debates over intellectual property boundaries in the rapidly evolving open-source AI landscape, emerged following a study published on GitHub on July 4.
The GitHub study, authored by an unnamed individual claiming to be a Korean student at the University of Costa Rica, alleged a striking resemblance between Huawei’s 72-billion-parameter Pangu Pro MoE model and Alibaba’s 14B Qwen-2.5. The study reportedly used a ‘model fingerprinting technique’ and cited a 0.927 correlation between the attention parameter distributions of the two models, suggesting potential non-independent development. Further evidence cited was leftover metadata in Huawei’s open-source code, which allegedly referenced ‘Copyright 2024 The Qwen team, Alibaba Group.’
Huawei responded on July 5, emphasizing that the Pangu Pro MoE is a foundational model developed ‘from scratch’ on its proprietary Ascend hardware platform. The company clarified that while the Pangu 72B model incorporates code from various open-source models, it was not trained on data from other companies. ‘We strictly adhere to open-source license requirements and clearly indicate copyright statements within the code,’ Huawei stated, adding that ‘referencing open-source industry practices is standard, and aligns with the collaborative spirit of the open-source community.’
The tech giant highlighted key innovations behind its Pangu Pro MoE, including a proprietary Grouped Mixture of Experts (MoGE) architecture designed to optimize load balancing in distributed training environments. Huawei stated that the model was trained using its Ascend GPUs and NPUs, which enhance overall training efficiency and address load balancing challenges in large-scale distributed training.
Huawei had open-sourced two Pangu AI models on June 30: the Pangu 7B dense model and the Pangu Pro MoE 72B hybrid expert model. This move was described as a strategic push to strengthen its AI ecosystem, allowing developers to access and customize the original code.
Despite the GitHub repository being taken down, the incident has sparked widespread discussion on social media platforms and Chinese developer forums like Zhihu. Some commenters have questioned the scientific validity of the fingerprinting methodology used in the plagiarism claim, arguing that parameter standard deviation alone is insufficient to determine model similarity.
This dispute underscores the complex legal and ethical gray areas prevalent in open-source AI development. While open-sourcing a model does not waive intellectual property rights, confusion persists regarding licensing obligations, attribution requirements, and commercialization restrictions. Experts caution that developers must clearly acknowledge original sources and comply with license terms, especially when adapting or commercializing open-source projects. This is not an isolated incident, as earlier this year, 01.AI’s Yi-34B faced criticism for borrowing architecture from Meta’s LLaMA, and Stanford’s Llama3-V was found to have repackaged Moonshot AI’s MiniCPM-Llama3-V 2.5.
Also Read:
- Baidu Pivots to Open-Source AI with Ernie 4.5 Release, Intensifying Global Competition
- Alibaba Cloud Commits $60 Million to Accelerate Global AI Adoption via Enhanced Partner Ecosystem
As the open-source model space becomes increasingly crowded with low technical barriers, allowing startups to quickly deploy services by wrapping existing models, the need for clearer licensing frameworks and enforcement mechanisms is becoming critical. Alibaba has yet to issue a public comment on the matter, while AI research teams from several major Chinese tech companies are closely monitoring the situation. The incident highlights the ongoing tension within the AI community regarding how to foster open-source innovation without compromising intellectual property protection or encouraging unchecked model replication.
