The White House has accused Chinese startup Moonshot AI of stealing from Anthropic to build Kimi K3, the open-weight model that knocked tech stocks around last week. Michael Kratsios, who runs the White House Office of Science and Technology Policy, posted on X on Wednesday that the US government has information showing Moonshot distilled Anthropic’s Fable model to create K3. The more damning part of his claim is what came next: Moonshot allegedly built an internal platform specifically to run this at scale, one that could switch between different ways of accessing US models so the activity stayed invisible.Kratsios stacked a second charge on top. He said Moonshot acquired servers fitted with Nvidia GB300 chips and used them in Thailand, likely for training. Those chips belong to the Blackwell generation, which Washington does not allow Chinese firms to buy anywhere in the world. Both accusations arrived a day after Treasury Secretary Scott Bessent went on Fox Business and said the US could sanction Chinese companies for what he called, without much hedging, theft. Moonshot has said nothing. Neither has the Chinese embassy in Washington.
Moonshot AI distilled Anthropic’s Fable model, says White House official Michael Kratsios
Distillation means training a smaller model on the outputs of a bigger one. It is ordinary practice, and Kratsios said so himself, calling legitimate distillation a vital part of the open innovation ecosystem. What he objects to is the scale and the concealment. His words: large-scale, covert industrial distillation aimed at stealing proprietary US technology and undermining American research is unacceptable.Anthropic has been making this case for months. In February it named Moonshot alongside DeepSeek and MiniMax, alleging Moonshot ran hundreds of fake accounts through roughly 3.4 million exchanges with Claude. In June it accused Alibaba of the largest distillation attack it had recorded. Sarah Heck, Anthropic’s head of public policy, thanked Kratsios on X and framed the practice as industrial espionage that ends up feeding foreign military and intelligence capability.
Kimi K3 is a 2.8 trillion-parameter open-weight model that beat US benchmarks
Moonshot showed K3 last week and called it the largest open-weight system in the world at 2.8 trillion parameters. On coding and general capability benchmarks it sits level with or above Anthropic and OpenAI’s frontier models. By Moonshot’s own accounting, only Claude Fable 5 and GPT-5.6 come out ahead overall. The full weights go up on July 27, free to download and modify, which is the part that actually worries American labs. Anthropic and OpenAI spent billions getting to this point. Moonshot is giving its version away.The timeline is also why the distillation claim has drawn scepticism. Fable came out about a week before K3, and Kratsios produced no evidence.
US sanctions on Chinese AI models are now on the table, says Scott Bessent
Bessent said the administration is finding watermarks from American models showing up in Chinese ones. He supports open source, he said, but open source is not open season on American IP. Axios has reported the White House is weighing a rule that would let US companies host Chinese models only if they guarantee security and accept liability for breaches, and that similar efforts were killed internally last year before national security hawks regained ground.The pushback is coming from inside the industry. Chamath Palihapitiya called it a China boogeyman used to protect the equity of a few thousand investors. Hugging Face CEO Clem Delangue was blunter: if distillation were the whole story, plenty of other countries would already have great open models. China’s research teams, he said, are simply better and more open than America’s.
