Senior figures in the White House have accused the Chinese AI developer Moonshot AI of distilling Anthropic’s most advanced model to develop its latest offering, as well as using black market Nvidia hardware for its training runs.
In a post on X Michael Krastios, director of the White House Office of Science and Technology Policy, formally accused Moonshot AI of distilling Claude Fable 5 in the development of Kimi K3, its open-source frontier model that competes directly with Fable 5 and OpenAI’s frontier model GPT-5.6 Sol.
Distillation is a process by which large language models (LLMs) are trained on the outputs of an exemplar model.
“To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of access to avoid detection,” he said.
“Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models.”
The Nvidia GB300, the firm’s highest-performance server rack based on its Grace Blackwell Ultra architecture, is not legally accessible in China. The US government has increasingly accused Chinese firms of circumventing import restrictions to access Nvidia hardware for AI training.
The US Department of Justice formally charged the co-founder of Super Micro Computer and two other people in March, alleging conspiracy to route $2.5 billion of Nvidia GPUs into China via Taiwan.
In recent weeks, Singaporean prosecutors filed charges against a suspect accused of defrauding server manufacturers, with authorities in the regional authorities pledging to crack down on firms in the region secretly routing advanced chips to China.
The Financial Times additionally reported that Nvidia has halved its list of approved Asian customers as the chip giant aims to stem the illicit flow of its advanced GPUs into China.
Sharing Krastios’s post on X, Jacob Helberg, under secretary of state for economic affairs in the Trump administration, wrote that Moonshot AI’s alleged activity amounted to “a heist of invaluable American Intellectual Property”.
“It is an assault on every economy that prizes the entrepreneur, rewards private capital, and relies on fair and honest competition,” he said.
However, some experts in the sector have already cast doubt on the administration’s claims. Chris Paxton, AI innovation lead at Agility Robotics and former research scientist at both Meta and Nvidia, questioned if it would have been possible for Moonshot AI to distil Fable 5 in the short time window between its 9 June release and the public launch of Kimi K3 on 16 July.
“How exactly could they have gotten enough data and trained a frontier model, much less tested it, in that one month,” he wrote on X. “If they have some way to distill a frontier model with all its capabilities from such minimal data that’s actually a huge accomplishment […] But unfortunately I just think this post is mostly wrong.”
These are not the first allegations of distillation that Moonshot AI has faced.
In February, Anthropic formally accused Moonshot AI of distilling its models using “hundreds of fraudulent accounts”. The US AI developer said Moonshot AI had logged 3.4 million exchanges with Claude, activity it believed the Chinese developer was carrying out to distil Claude’s coding, reasoning, and computer vision capabilities.
Neither the US government nor Anthropic have yet provided any evidence of the claims. If the model was distilled from Fable 5 as alleged by US officials, this could be proven through forensic analysis of Kimi K3’s weights, the numerical parameters within the LLM’s neural network that determine its performance and behaviour, which Moonshot AI is set to release on July 27.
