Huawei Denies Pangu Model Copied Alibaba’s Qwen * SciTechTimes.com

Huawei’s Noah Ark Lab refutes claims its Pangu Pro Moe model was “upcycled” from Alibaba’s Qwen 2.5 14B, insisting on independent development.

Allegations Trigger Industry Debate

Late last week, an entity named HonestAGI published a GitHub paper alleging that Huawei’s Pangu Pro Moe large language model bore “extraordinary correlation” with Alibaba’s Qwen 2.5-14B. The report asserted that statistical “fingerprints” across attention matrices pointed to an “upcycling” of Qwen rather than training from scratch. Authors further claimed potential copyright infringement, fabrication in technical documentation, and exaggerated assertions of Huawei’s training investments. Reactions rippled through AI circles and Chinese tech media, sparking fresh scrutiny of model provenance in the race for generative AI supremacy.

Noah Ark Lab’s Firm Rebuttal

Noah Ark Lab, Huawei’s flagship AI research division, issued a swift statement denying any derivative work. Team representatives emphasized key architectural innovations and unique technical features developed in-house on Huawei’s proprietary Ascend chips. Adherence to open-source licensing was also stressed, with the lab confirming that any third-party code used complied fully with license requirements. The statement did not specify which open-source frameworks served as references, but insisted that Pangu Pro Moe was “independently developed and trained.”

Stakes in China’s AI Ecosystem

Competition between China’s AI heavyweights has intensified since DeepSeek’s R1 open-source debut early this year. Alibaba’s Qwen series, designed for consumer and chatbot applications, contrasts with Pangu’s focus on enterprise, government, finance, and manufacturing use cases. As U.S. export controls limit access to Nvidia GPUs, domestic chipsets like Huawei’s Ascend become vital. Industry observers warn that disputes over model origins could accelerate calls for standardized provenance verification tools, ensuring fair competition and legal compliance across global AI development.

Interesting Insight

Model fingerprinting, the technique cited by HonestAGI, analyzes the standard deviation patterns in a model’s attention parameter matrices (Q, K, V, O). This statistical “signature” is believed to persist even after continued training, making it possible to trace a model’s lineage despite efforts to mask its origins. Independent researchers deployed similar methods in 2023 to flag suspected reuse of Meta’s LLaMA code in various open-source forks. As Chinese tech giants accelerate open-source releases to sidestep hardware sanctions, robust provenance tracking will become a critical safeguard against inadvertent code reuse and potential IP disputes.