Did China Steal From Anthropic? Experts Doubt White House Claims

A senior White House official has publicly accused the Chinese company Moonshot AI of building its new Kimi K3 model by copying American artificial intelligence from Anthropic. This marks the first time a U.S. official at this level has named both a specific Chinese laboratory and a specific stolen model. However, a number of researchers in the field immediately objected that the entire accusation does not make technical sense.

What Washington is accusing Moonshot of

Michael Kratsios, who heads the White House Office of Science and Technology Policy and also serves as President Trump’s science adviser, wrote on X that the U.S. administration has information indicating that Moonshot distilled Anthropic’s Fable model to create its K3 model. According to Kratsios, the Chinese company built a sophisticated internal platform for large-scale distillation of U.S. models, enabling it to switch quickly between different access methods to avoid detection.

Distillation is a process in which a smaller model learns to imitate the outputs of a larger and more capable model. The stronger model serves as the teacher and the smaller one as the student. The student does not learn merely from what is right and wrong, but also observes the teacher’s entire probability distribution of responses—for example, that one word has a ninety-percent probability and another an eight-percent probability. It is precisely from these subtle differences that it infers how the teacher “thinks.”

Kratsios himself distinguished between two forms of this technique. He described ordinary distillation, which creates smaller and more efficient models and is useful for open development, as legitimate. By contrast, he characterized large-scale covert industrial distillation aimed at stealing American technology and undermining U.S. research as unacceptable. However, he did not publish any technical documentation or other evidence to support the claim.

Back in February, Anthropic publicly accused Moonshot, along with DeepSeek and MiniMax, of systematically distilling its models. At the time, it claimed to have traced roughly 3.4 million conversations with Claude to Moonshot and reported more than sixteen million interactions in total through approximately 24,000 fraudulent accounts, allegedly violating its terms of use and regional access restrictions.

Banned chips and sanctions evasion

The second part of the accusation concerns hardware. Kratsios claims that Moonshot obtained servers equipped with Nvidia GB300 chips and also used machines with GB300 chips located in Thailand, probably to train its models. The United States prohibits the export of these Blackwell-generation chips to China.

The rhetoric had already intensified a day earlier. Treasury Secretary Scott Bessent warned that many Chinese models contain watermarks from American language models, which he called unacceptable. He also issued a stronger statement saying that open source does not mean free rein over American intellectual property. He threatened companies engaging in large-scale distillation attacks with sanctions and placement on the so-called Entity List. This would cut Moonshot off from U.S. hardware, software, and cloud infrastructure, much as Washington did to Huawei in 2019. Deputy Secretary of State Jacob Helberg then went so far as to call Moonshot’s alleged distillation the theft of invaluable American intellectual property.

Sam Bresnick of Georgetown’s Center for Security and Emerging Technology notes that a black market exists for advanced chips. In May, the founder of U.S. server manufacturer Supermicro was charged with smuggling cutting-edge chips into China. Bresnick advocates requiring data centers around the world to know their customers. When a company trains massive models on state-of-the-art hardware, he believes there should be a mechanism for reporting who it is and what it is doing. The Biden administration’s Commerce Department proposed similar rules for data centers back in 2024, but they have apparently gone nowhere under Trump.

Moonshot rejects the accusation

The company itself has officially remained silent, but one of its employees has spoken out. Moonshot employee Randy Xian countered on X with a timeline. According to him, Fable was released on July 1 and Kimi K3 appeared on July 15. “We trained an entirely new state-of-the-art model in just 15 DAYS. A Guinness World Record,” he wrote sarcastically. He also pointed out that Fable had been offline for roughly ten days, so he could not see how they could have distilled anything in such a short time. He added that he almost viewed the entire accusation as a compliment, because in his view the Americans were merely demonstrating what a good job Moonshot had done.

It is worth recalling that Fable 5 became unavailable because of U.S. export restrictions and returned to the public precisely on July 1. Kimi K3 was released as an open model with approximately 2.8 trillion parameters, making it the largest freely downloadable language model ever. Moonshot is expected to release the complete weights at the end of July, which could reveal a great deal.

Experts doubt the distillation claim

And it is precisely this fifteen-day gap that most expert reactions focus on. Braden Hancock of the Laude Institute and co-founder of Snorkel AI told TechCrunch that a model this powerful could not be produced purely through distillation in such a short time. There could not have been enough time to obtain that much data, train the model, and release it within two weeks.

Others have made similar arguments. Elie Bakouch of Prime Intellect pointed out that only fifteen days elapsed between the lifting of the Fable 5 ban and the release of Kimi K3, and that claiming K3’s performance came from distilling Fable makes no technical sense to him, even if some distillation did occur. Dean Ball, who heads strategic planning at OpenAI, likewise does not believe the model’s performance can be explained by distillation.

Distillation becomes less effective as Chinese models approach the cutting edge and as training shifts toward reinforcement learning. If distillation alone were enough, they argue, anyone could catch up with K3 or the GLM model simply by using their data. But that is not happening, nor will it happen merely through fine-tuning on finished responses, known as supervised fine-tuning. It is precisely through this type of fine-tuning that a model “picks up mannerisms,” which, incidentally, is also why a model created by someone else may occasionally claim that it is Claude.

According to analysts, replicating capabilities similar to Fable would require reinforcement learning, which is in an entirely different league. Large reinforcement-learning training runs require tens of millions of agents and massive infrastructure. If someone wanted to conduct this through the interface of a leading laboratory, it would be prohibitively expensive, they would encounter problems with the speed of those models, and it would most likely fail to improve performance anyway.

Distillation itself is certainly not a Chinese specialty. Elon Musk testified in court this year that his company had distilled OpenAI’s models to develop Grok, describing it as standard practice in the industry. Moreover, the boundary between distillation and the creation of synthetic data is often quite blurred. Hancock adds that Americans underestimate the technical sophistication of Chinese teams. One of Moonshot’s founders was a doctoral student at Carnegie Mellon; these are genuine researchers and engineers doing solid work. Even if American models stopped advancing, he believes Chinese progress would slow down but would not stop. The Chinese are certainly not merely riding on others’ coattails.

Sources: businessinsider.com, hongkongfp.com and cnn.com