Michael Kratsios, the White House science and technology adviser, accused China's Moonshot AI of distilling Anthropic's Fable model to build its 2.8 trillion parameter Kimi K3.
Key Points:
- The White House says Moonshot AI ran a hidden platform to pull capabilities out of US frontier models.
- Officials also allege the company reached restricted Nvidia GB300 servers in Thailand.
- Treasury has raised the prospect of sanctions and Entity List designations against Chinese AI firms.
Kratsios Details Moonshot AI Distillation Claims
Kratsios, who directs the White House Office of Science and Technology Policy, wrote on X on Wednesday that Moonshot built an internal platform to run large scale distillation against American models.
The system let the company switch between access methods and avoid detection, he said.
He also alleged that Moonshot acquired servers equipped with Nvidia's GB300 chips and tapped GB300 hardware in Thailand. Washington restricted exports of the chipmaker's most advanced processors in 2022 over concerns about Chinese military use. Governments have since moved against the smuggling routes that carry them.
Distillation trains a weaker model on answers drawn from a stronger one, a practice Kratsios called vital when done in the open and unacceptable when hidden.
Also Read: Anthropic Chased An $11B Robot Startup Before The Denial Went Public
Bessent And Lieu Split Over Sanctions Threat
Treasury Secretary Scott Bessent told Fox Business on Tuesday that the government can sanction foreign models built on stolen American work. Officials keep finding watermarks from US large language models inside Chinese systems, he said.
Bessent added a day later that sanctions and Entity List designations would be on the table for firms running industrial scale distillation attacks. Representative Ted Lieu questioned that position while the administration still clears sales of high performance AI chips to China. Kratsios gave no detail on how the government learned that K3 had been distilled.
Kimi K3 Weights Follow Anthropic Complaints
Moonshot launched Kimi K3 on Jul. 16 and has not answered the latest accusations, which Beijing called groundless when Anthropic raised similar ones earlier this year.
The model carries 2.8 trillion parameters, the largest open weight release to date, and its weights go public on Jul. 27.
Anthropic first accused Moonshot in February, alongside DeepSeek and MiniMax, of using roughly 24,000 fake accounts to generate about 16 million interactions with Claude.
In June the company said Alibaba's Qwen unit ran nearly 25,000 fraudulent accounts and produced more than 28.8 million interactions. It asked Congress to penalize the firms behind those campaigns.
Read Next: Hulu's Charles Manson Documentary Uncovers His Unknown Granddaughter As Prediction Markets Watch