China’s AI unity fractures as Huawei faces model theft allegations from the Alibaba camp

Huawei’s AI research division has rejected claims that its Pangu Pro large language model copied elements from an Alibaba model, marking a significant escalation in China’s AI ecosystem as tech giants abandon their collaborative approach in favor of bitter public disputes.

The telecommunications giant’s Noah Ark Lab issued a denial Saturday, after an entity called HonestAGI published a technical analysis claiming Huawei’s Pangu Pro Mixture of Experts (MoE) model showed extraordinary correlation with Alibaba’s Qwen 2.5 14B model, reported Reuters. The analysis alleged the model was derived through “upcycling” rather than being trained from scratch.

The public confrontation represents a dramatic shift from China’s previous unity in challenging Western AI dominance. Industry analysts say the infighting could undermine China’s ability to present a consolidated front against US-led competitors like OpenAI, Google DeepMind, and Anthropic.

HonestAGI’s GitHub analysis claimed a correlation coefficient of 0.927 between the two models, using what it called “model fingerprinting” to identify patterns that supposedly revealed one model’s derivation from another.

Noah Ark Lab responded that its model was “not based on incremental training of other manufacturers’ models” and featured “key innovations in architecture design and technical features.” The company emphasized that Pangu Pro was the first large-scale model built entirely on Huawei’s Ascend chips, the report added.

“This dispute actually points to changing dynamics of the Chinese AI ecosystem’s speed of maturity and pressure to remain relevant and compete to foster innovation faster than the traditional collaborative approach, which we have seen,” said Neil Shah, VP for research and partner at Counterpoint Research.

Competition reaches fever pitch

The controversy escalated when an alleged Huawei insider posted detailed accusations about systematic model copying within the company. The anonymous whistleblower, claiming to be a Pangu team member, accused leadership of “cloning” both Alibaba’s Qwen and startup DeepSeek’s models while presenting them as original work.

“They had ‘cloned’ Qwen‑1.5 (110B), wrapped it in extra layers and changes — creating a pseudo‑135B ‘V2’ model,” the whistleblower wrote in the paper. “This rebranded model, with code still named ‘Qwen,’ was rolled out to clients.”

The allegations couldn’t be independently verified, and the whistleblower’s identity remains unknown.

The dispute comes as Chinese AI companies scramble after DeepSeek’s breakthrough R1 model release in January stunned Silicon Valley with its low-cost, high-performance approach. Alibaba rushed out its Qwen 2.5-Max model just weeks later, claiming superior performance across multiple benchmarks.

“What once was a state-aligned innovation drive is now being reshaped by market-led competition, where speed-to-scale often overrides transparency,” said Sanchit Vir Gogia, chief analyst and CEO at Greyhound Research.

Trust deficit emerges

This development has raised uncomfortable questions about credibility on all sides. Technical analysis of HonestAGI’s methodology revealed potential flaws, with researchers finding similar correlation patterns between unrelated models using the same fingerprinting technique. Critics also discovered fabricated references to non-existent research in HonestAGI’s paper.

“Also, this is a double-edged sword for China’s strategy to drive openness of the models where there could be potential derivations of the best models out there,” Shah added. “We have seen this happen with OpenAI-DeepSeek as well.”

The dispute highlights broader challenges facing the AI industry as development costs soar and model reuse becomes common. Vershita Srivastava, practice director at Everest Group, said the sector needs better tools to handle such controversies.

“The industry must adopt a comprehensive framework that includes advanced fingerprinting and watermarking techniques that can reliably trace model lineage,” Srivastava said.

The public nature of this dispute marks a turning point for China’s AI sector, which previously maintained at least a veneer of collaboration.

Gogia warned that the infighting could have lasting consequences beyond China’s borders. “This episode underscores that Chinese vendors are now operating under public scrutiny, and any erosion of trust could have lasting geopolitical and commercial consequences,” he said. The controversy may force enterprise buyers, especially in Southeast Asia and the Middle East, to reevaluate partnerships with Chinese AI providers.

The allegations have also exposed what Gogia calls the “growing inadequacy of conventional IP frameworks when applied to LLMs.” Parameter-level fingerprinting techniques offer promise but remain scientifically contested and legally untested.

Market divide

The feud highlights how China’s AI leaders target different markets while chasing the same prize. Alibaba’s Qwen family focuses on consumer applications with ChatGPT-like services and has been downloaded more than 40 million times since going open-source. Huawei’s Pangu models target enterprise clients in government, finance, and manufacturing.

Despite entering the large language model arena early with Pangu’s 2021 debut, Huawei has struggled to keep pace with rivals. The company open-sourced its Pangu Pro MoE models in June, hoping to boost adoption through free developer access.

The latest controversy underscores the urgent need for industry-wide standards. “Without agreed-upon definitions of derivation — particularly in models trained on shared corpora — vendors face an unclear compliance landscape,” Gogia noted. “This ambiguity creates space for weaponized accusations and erodes open-source collaboration.”

Srivastava emphasized the need for legal frameworks, saying it’s “imperative to establish clear definitions for derivative models and implement nuanced licensing frameworks that support responsible reuse, enforce appropriate attribution, and uphold usage restrictions.”

How this controversy resolves will set important precedents for intellectual property disputes in an increasingly competitive AI landscape. The success of nimble operations like DeepSeek has upended assumptions about what it takes to build cutting-edge AI, making bloated bureaucracies look more like liabilities than advantages.

Alibaba did not immediately respond to requests for comment.