The White House publicly called out Chinese AI startup Moonshot AI on July 22, accusing it of building its newly launched Kimi K3 model partly through the unauthorized mass extraction of capabilities from Anthropic's Fable system—escalating Washington's campaign against what it describes as industrial-scale intellectual property theft in China's AI sector.
A Celebrated Launch Followed Quickly by Serious Accusations
Moonshot AI released Kimi K3 on July 16, billing it as one of the world's largest open-source AI systems, with a reported 2.8 trillion parameters. Benchmark results showing the model closing performance gaps with Anthropic's Fable 5 and OpenAI's GPT-5.6 drew wide attention across the global AI research community.
That attention turned hostile less than a week later. Michael Kratsios, director of the White House Office of Science and Technology Policy, posted a public accusation on X alleging that Moonshot AI had conducted large-scale "distillation" of Anthropic's Fable model during Kimi K3's development. Kratsios also alleged the company had obtained Nvidia GB300 Grace Blackwell servers—hardware subject to US export controls—to train the model, with officials suggesting the procurement may have been routed through overseas infrastructure in countries including Thailand.
Why Distillation Has Become Washington's Preferred Legal Weapon
Knowledge distillation is, in itself, a legitimate and widely used machine learning technique. The process trains a smaller "student" model on the outputs of a larger "teacher" model, allowing the student to approximate the teacher's performance at significantly lower computational cost. Academic and enterprise use of distillation is routine.
The controversy turns on authorization. When a company routes millions of API queries through another company's proprietary closed-source model—without permission—and uses those outputs to train a competing product, US officials argue this crosses into terms-of-service violation and potential intellectual property infringement. The White House has characterized Moonshot AI's alleged behavior as going far beyond ordinary model optimization, framing it instead as "industrial-scale capability extraction."
Anthropic Flagged Moonshot AI in an Earlier Complaint
This is not the first time Moonshot AI has faced these accusations. Earlier this year, Anthropic said that Chinese AI firms—including Moonshot AI, DeepSeek, and MiniMax—had used more than 16 million unauthorized interactions with its Claude models to systematically collect training data. Alibaba was also named in similar allegations. Anthropic said the activity appeared highly organized and that it had since strengthened its detection and prevention mechanisms.
The pattern has drawn increasingly pointed responses from Washington. Treasury Secretary Scott Bessent has warned that Chinese companies crossing intellectual property lines risk placement on sanctions or entity lists. A White House memo issued in April framed the broader practice as a national-level threat aimed at systematically undermining US research and development capabilities. Some American AI companies have already begun restricting Chinese individual developers from accessing their models through direct connections.
Beijing Pushes Back as Moonshot Moves to Open-Source K3
China's Foreign Ministry spokesperson Lin Jian rejected the accusations, saying Beijing opposes the politicization of technology and trade issues and that China's AI development is grounded in principles of "consultation, co-building, and sharing"—reflecting, in Beijing's telling, independent progress and legitimate international collaboration.
Moonshot AI announced plans to open-source Kimi K3's model weights on July 27, making them freely available for developers worldwide to download and customize. The move may be partly intended to project transparency, though whether it will ease or intensify US scrutiny remains to be seen as Washington deepens its investigation into alleged Chinese AI capability theft.
