Whistleblower Exposes Alleged Fraud and Misconduct in Huawei's Flagship AI Program
Inside the Broken Dreams of China's AI Ambitions
A core member of Huawei's prestigious Noah's Ark Lab has published detailed allegations claiming widespread misconduct, plagiarism, and academic dishonesty in the development of the company's flagship Pangu large language models.
The anonymous whistleblower's account, published on GitHub on July 9, 2025, offers an unprecedented glimpse into what they describe as "the heart of darkness" within one of China's most celebrated tech giants. According to the letter, Huawei's celebrated achievements in AI may be built on falsified results, stolen intellectual property, and a culture that "punishes integrity while rewarding deceit."
"Shelling" Competitors' Models: The Alleged Grand Deception
At the center of the allegations is a practice described as "shelling" – where Huawei allegedly took competitor models, made superficial changes, and presented them as proprietary developments. The whistleblower claims that under pressure to deliver results, a team led by "Wang Yunhe's Small Model Lab" repackaged Alibaba's Qwen-110B model, tweaked it slightly, and renamed it "135B V2."
Internal analysis reportedly revealed damning evidence: mismatched architecture, identical parameter distributions to Qwen, and source code still containing "Qwen" names. The whistleblower alleges this model was deployed to downstream clients and celebrated internally, despite many team members being "horrified" by the deception.
"The model wasn't just similar – it was literally Qwen with a Huawei badge slapped on it," one AI researcher familiar with the situation told this reporter, speaking on condition of anonymity due to fear of reprisal. "Anyone with the technical knowledge to examine the architecture could see the truth."
Benchmark Fabrication: The Perfect Scores That Couldn't Be
Perhaps most damaging are allegations surrounding Huawei's published benchmarks for Pangu Ultra. The whistleblower claims the model report showed a mathematically impossible 100% accuracy on the ARC-Easy benchmark – a result immediately flagged as "unrealistic or fabricated" by external experts.
Former Huawei Noah's Ark Lab engineers provided a troubling explanation: instead of testing on full datasets (approximately 5,200 questions for ARC-Easy), the team evaluated only 100-sample subsets used for internal quick checks. These partial results were allegedly included in the final public report under time pressure, creating the illusion of perfect scores.
Additionally, benchmarks like RACE were evaluated using simplified methods that inflated scores by up to 40 points compared to traditional perplexity-based methods used by competitors. These inflated results were then directly compared against other models' scores evaluated using stricter methods – a misleading comparison that violated academic norms.
The Whistleblower's Anguish: "Blood, Sweat, and Sacrifice"
The letter reveals a deeply personal dimension to the scandal. The author describes years of grueling work by dedicated engineers who believed in Huawei's mission to build domestic alternatives to NVIDIA's AI hardware. The team reportedly trained progressively larger models on Huawei's Ascend NPUs, facing significant technical challenges.
Unlike the allegedly plagiarized 135B V2, the whistleblower claims the 135B V3 model (Pangu Ultra) was "genuinely trained from scratch" by their team, using a refined tokenizer and improved training pipeline. This model – described as the "true product of blood, sweat, and sacrifice" – delivered competitive performance with "clean training and no loss spikes," a rare feat in large model training.
"I can accept bad results. I can't accept dumb ones," stated one former engineer identified as "Blealtan," a PhD from Tsinghua previously responsible for MoE infrastructure at Noah's Ark Lab.
A Culture of Suppression and Bureaucratic Rot
When engineers attempted to correct the flawed benchmark reports, they were allegedly blocked by senior leadership who "feared public backlash and preferred avoiding any amendments that might admit mistakes." The whistleblower describes a toxic environment where honest teams burned out or left while dishonest actors gained recognition and resources.
Author lists for research papers were reportedly "curated by management, not based on actual contributions," with some people discovering their names had been added or removed only after reports were made public. The uploading of research to platforms like arXiv was allegedly done by non-technical personnel with minimal input from core team members.
The Real Stakes: China's AI Ambitions in the Balance
The allegations come at a critical juncture for China's AI sector, which has been racing to close the gap with U.S. competitors like OpenAI and Anthropic. Huawei's Pangu models are central to the country's ambitions for technological self-sufficiency amid ongoing U.S. sanctions.
Industry experts suggest Huawei's alleged misconduct may have been motivated by more than pride. Some argue the goal was to use better-performing external models to falsely prove Huawei's Ascend chips are just as capable as NVIDIA's for training top-tier LLMs – potentially helping Huawei sell Pangu-integrated hardware to government and military clients.
Investor Outlook: Navigating the Fallout
For investors watching China's AI race, the allegations raise significant concerns about the true state of domestic AI capabilities. Market analysts suggest several potential implications:
First, companies developing genuinely innovative AI technologies with transparent research practices may emerge as more reliable long-term investments. Firms like DeepSeek and Baidu, which have emphasized open-source approaches and verifiable benchmarks, could benefit from increased scrutiny of AI claims.
Second, the hardware angle deserves particular attention. If Huawei's Ascend chips truly struggle with large-scale AI training compared to NVIDIA's offerings, supply chain companies supporting NVIDIA could maintain their competitive advantage longer than previously anticipated.
Finally, the scandal may accelerate regulatory oversight of AI benchmark reporting in China, potentially creating compliance challenges but also opportunities for companies offering third-party verification services.
"What we're witnessing could be a watershed moment for China's AI sector," noted one technology investment strategist. "The market will likely reward transparency and punish opacity going forward."
Disclaimer: This analysis reflects current market conditions and established indicators only. Past performance does not guarantee future results. Readers should consult financial advisors for personalized investment guidance.
As Huawei has yet to officially respond to these allegations, the full impact remains uncertain. What is clear, however, is that beneath the glossy announcements and impressive benchmarks of China's AI revolution, there may be far more complexity – and controversy – than previously understood.