The US government accuses the Chinese company Moonshot AI of developing its model Kimi K3 through large-scale, covert distillation of Anthropic’s Fable. This was stated by science advisor Michael Kratsios on July 22 on the platform X. Treasury Secretary Scott Bessent threatens sanctions and entries on the Entity List in the event of confirmed theft.
Kratsios describes obfuscation platform with multiple access points
According to Kratsios, Moonshot developed internal software that automatically switched between multiple access paths to US models to obscure systematic distillation. This is the first time a high-ranking US government official has publicly accused a specific Chinese company of model distillation. US Deputy Secretary of State Jacob Helberg also commented sharply on the case.
Kimi K3 was created this way: an open 2.8 trillion parameter model with one million token context, which Moonshot released on July 15 – just over two weeks after Anthropic relaunched its model Fable on July 1, as beckmann.ai reported at the launch.
Additionally, Moonshot is said to have used servers with Nvidia’s GB300 chips, which fall under US export controls and are officially not allowed to be sold to China. According to Kratsios, the company sourced the hardware through Thailand. Treasury Secretary Scott Bessent stated that the approach is not an open invitation to steal American intellectual property. In cases of industrial, covert technology theft, sanctions and entries on the Entity List are possible. A congressional committee has been investigating alleged distillation attacks by Chinese providers since April.
Moonshot denies the allegations, experts doubt the timeline
A Moonshot employee responded via the platform X to the allegations: “We trained a completely new frontier model in just fifteen days.” He referred to the time gap between Fable’s relaunch on July 1 and Kimi K3’s release on July 15. An official Moonshot statement was not provided, and several news outlets received no response to inquiries.
AI researcher Braden Hancock from the Laude Institute believes, according to TechCrunch, that it is unlikely such a strong model could emerge in just two weeks solely through distillation. Nathan Lambert from the Allen Institute for AI adds that distillation is becoming increasingly ineffective for Chinese providers with a growing technical gap from the top. Elaborate reinforcement learning over millions of agents is hardly financially feasible through mere API requests.
Anthropic had already stated in February that several Chinese providers – including Moonshot, DeepSeek, and MiniMax – had attempted to extract the capabilities of Claude models through fake accounts and proxy services. Moonshot reportedly made over 3.4 million requests through hundreds of accounts, focusing on agentic reasoning and computer usage. Whether the same infrastructure was also used for Kimi K3 remains open in Anthropic’s current case.
It will be crucial whether the US government presents technical evidence beyond the previous claims – without public evidence, the case remains a he-said-she-said situation between Washington and Beijing. For Moonshot, more is at stake than just reputation: the company is reportedly preparing for an IPO in Hong Kong with a targeted valuation of 30 billion dollars. Sanctions or an Entity List entry could significantly complicate this plan, as well as access to further urgently needed computing power.
