appConfigEntityId

bodyType

customContents

entityId

entityUuid

isVideo

articlePageQuery

article

articleSchemaContent

content

hooksArticlePageAdvertisementConfigArticle

hooksArticlePageFloatingButtonConfigArticle

appConfig

filter

json

liveArticle

Vincent Chow

The Trump administration has accused Chinese start-up Moonshot AI of covertly extracting capabilities from leading US artificial intelligence models and obtaining restricted Nvidia chips abroad, escalating Washington’s scrutiny of the company following the release of its powerful Kimi K3 model.
Michael Kratsios, the White House science and technology adviser, alleged on Wednesday that the Beijing-based company had targeted Anthropic’s most powerful model, Claude Fable 5, through large-scale “distillation” – a technique that uses the outputs of a more capable model to train or improve another system.
Kratsios did not present evidence that outputs from Fable 5 had been incorporated into Kimi K3; however, the timing of the two models’ releases prompted disagreement among AI researchers about whether Moonshot could have used the US model to train its latest system.
Kimi K3, released last week, has attracted attention for its performance in coding, reasoning and other complex tasks.
The open-weight model has 2.8 trillion parameters, making it one of the largest AI systems released by a Chinese developer.


The accusations reflect Washington’s growing concern that Chinese developers may be circumventing two pillars of US efforts to preserve its AI advantage: restrictions on access to proprietary American models and controls on exports of advanced semiconductors.
In his social media post, Kratsios said that Moonshot had “developed a sophisticated internal platform to conduct large-scale distillation against US models, allowing them to quickly switch between multiple methods of access to avoid detection”.
“Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem,” Kratsios said.
“However, large-scale, covert industrial distillation aimed at stealing proprietary US technology and undermining American research is unacceptable.”
Distillation itself is a widely used industry practice and is not inherently illicit.
US companies contend, however, that Chinese developers engage in a form of “adversarial” distillation by circumventing geographic restrictions and terms of service that bar the use of their models in China.
In February, Anthropic accused Moonshot of using hundreds of fraudulent accounts, purportedly created by senior company personnel, to generate millions of exchanges with its models as part of a broader effort to extract its capabilities.
The US company said at the time that Moonshot and two other Chinese AI companies, DeepSeek and MiniMax, had collectively used about 24,000 fraudulent accounts to generate more than 16 million exchanges with Claude.

However, the company has yet to show that outputs from Fable 5 had been used specifically to develop Kimi K3.
Moonshot did not immediately respond to a request for comment.
AI researchers question whether Moonshot had time to extract Fable 5 outputs
The lack of publicly disclosed evidence linking Fable 5 directly to Kimi K3 has become a focus of debate among AI researchers.
Fable 5 was released publicly only last month, raising questions about whether Moonshot would have had enough time to extract its outputs and incorporate them into K3 before releasing the Chinese model last week.
Nathan Lambert, an AI expert and author of the AI newsletter Interconnects, expressed doubts that Moonshot could have distilled Fable 5, which has some of the industry’s strictest restrictions on model use and was suddenly suspended by the Trump administration last month over national security concerns.
“I personally am sceptical of evidence that Fable was distilled to train K3 directly,” Lambert wrote on X on Wednesday.
Ben Hayum, a research assistant at the Washington think tank Centre for a New American Security who specialises in AI distillation, said the Chinese AI industry tends to release its models more quickly after model training is completed than the US industry, which means that Kimi K3 might have incorporated outputs from Fable 5.
“If China releases models almost immediately after training finishes, then there would definitely be enough time for them to incorporate distilled data from Fable 5,” he said.


Distillation has emerged as a major point of contention in US-China AI competition. Besides Anthropic, other US companies and officials have also accused leading Chinese AI developers, including Alibaba Group Holding and Zhipu AI, of distilling US models. Alibaba owns the South China Morning Post.
US Treasury Secretary Scott Bessent said on Tuesday that Chinese companies deemed to have engaged in adversarial distillation could be sanctioned, while a new bipartisan amendment proposed in the Senate to the National Defence Authorisation Act, the annual defence policy bill, would codify such sanctions, including export controls and blacklisting through the US Entity List.
Chinese AI companies have generally not publicly disclosed whether or how they use outputs from American models to develop their own systems.
In a statement, the Chinese embassy in Washington spokesperson Liu Chang said that allegations against Chinese companies of illicit distillation are “entirely unfounded”.
“China has consistently attached great importance to the protection of intellectual property and is committed to safeguarding, in a fair and impartial manner, the lawful rights and interests of both domestic and foreign IP holders,” said Liu.
“Relevant individuals in the US should respect the facts, discard prejudice, and stop smearing and discrediting China’s achievements in the development of its AI industry.”
White House alleges Moonshot also obtained Nvidia’s advanced Blackwell GB300 chips
Kratsios separately alleged that Moonshot had obtained Nvidia’s advanced Blackwell GB300 chips and deployed them in Thailand, “likely to train its AI models”. He did not disclose the evidence supporting that assertion.
The claim is potentially significant because US export controls prohibit the direct sale of Nvidia’s most advanced Blackwell processors to China.
Washington has increasingly focused on the possibility that Chinese companies could gain access to restricted computing power through data centres, subsidiaries or other arrangements in third countries.
US officials have previously raised concerns that Chinese companies could access advanced Nvidia chips through operations in Southeast Asian countries, including Malaysia.
The US Department of Commerce issued guidance in May warning companies that using advanced US chips to train Chinese AI models could violate export controls, even when the processors were located outside China.
Nvidia did not immediately respond to a request for comment on Kratsios’ allegation concerning the chips in Thailand.
A spokesperson for the US Bureau of Industry and Security, which oversees US export controls, said that Moonshot’s reported access to Nvidia chips “is under investigation”, and “any wrongdoing will be subject to strong enforcement”.


Former WTO head Pascal Lamy on the EU’s China choice: get a rebalance, or go protectionist

US announces AI dialogue with China, launches Board of Trade

China elevates Li Chenggang to top international trade role ahead of Xi’s US trip

Boeing’s 200-jet China deal could hinge on Trump-Xi talks: industry insiders

US-China trade war

Trump tech official accuses China’s Moonshot AI of stealing from Anthropic

The Trump administration claims that outputs from Anthropic’s most powerful model, Claude Fable 5, have been incorporated into Moonshot AI's new Kimi K3 model. Photo: Shutterstock

White House science and technology adviser Michael Kratsios has accused Chinese start-up Moonshot AI of covertly copying Anthropic’s most advanced model to build its Kimi K3. Photo: Reuters

As US-Chinese AI model gap narrows, what next for Washington?

China’s latest AI model advancements have forced Washington to ponder its next move in the global artificial intelligence race. Image: Shutterstock

People walk past the Moonshot booth promoting its Kimi K3 AI model during the World Artificial Intelligence Conference in Shanghai on July 17. Photo: Reuters

China’s Kimi K3 fuels fears safety curbs are holding back US AI

Why China’s open-weight AI model Kimi K3 is sparking anxiety in Silicon Valley

People visit the Moonshot AI stand, featuring the Kimi K3 model, during the World Artificial Intelligence Conference in Shanghai on July 18. Photo: AFP

Clash of the tech titans: China’s quiet industry vs loud US innovation

The strengths of each model lie in its strategic priorities, as do the weaknesses. Photo: Shutterstock


News

Diplomacy

Economy, Trade & Business

China

Michael Kratsios alleges Beijing-based company targeted Anthropic’s most powerful model, Claude Fable 5, through large-scale ‘distillation’


Michael Kratsios alleges Beijing-based company targeted Anthropic’s most powerful model, Claude Fable 5, through large-scale ‘distillation’.


Trump tech official accuses China’s Moonshot AI of stealing from Anthropic

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}Michael Kratsios alleges Beijing-based company targeted Anthropic’s most powerful model, Claude Fable 5, through large-scale ‘distillation’

Michael Kratsios alleges Beijing-based company targeted Anthropic’s most powerful model, Claude Fable 5, through large-scale ‘distillation’