White House Accuses China's Moonshot AI of Covertly Copying Anthropic's Fable Model for Kimi K3
Key Takeaways
- •White House official Michael Kratsios alleged that Moonshot AI operated a purpose-built internal platform to conduct large-scale distillation of Anthropic's Fable model for its Kimi K3 system.
- •Kratsios stated that Moonshot used NVIDIA GB300 servers located in Thailand, chips prohibited from reaching China under U.S. export controls.
- •Anthropic's February report traced more than 3.4 million Claude conversations to Moonshot through hundreds of fabricated accounts, some matching profiles of Moonshot senior staff.
- •Kimi K3, at 2.8 trillion parameters the largest open-weight model released to date, is scheduled to have its full model weights made publicly available on July 27.
- •The allegations add to a series of U.S.–China AI copying disputes, including a January 2025 claim that DeepSeek copied OpenAI's models, ahead of bilateral AI talks set for September.
A senior White House technology official has accused Chinese AI startup Moonshot AI, one of China's most prominent AI ventures, of covertly copying Anthropic's Fable model to build its new Kimi K3 AI system, intensifying the ongoing U.S.–China technology dispute just as Moonshot prepares to release the model's full weights to the public.
Kratsios: "Undermining American research is unacceptable"
Michael Kratsios, director of the White House Office of Science and Technology Policy, made the allegations in a post on X on July 22, 2026. He stated that his office possessed information indicating Moonshot operated a purpose-built internal platform to conduct large-scale distillation against U.S. models, rotating access routes repeatedly to avoid detection.
Kratsios also said the operation relied on NVIDIA GB300 servers located in Thailand — chips that rank among the manufacturer's most powerful and that U.S. export controls prohibit from reaching China. Thailand is not subject to the same export restrictions that bar those advanced chips from reaching Chinese soil.
"Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable," Kratsios wrote.
We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 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…
— Director Michael Kratsios (@mkratsios47), July 22, 2026
Distillation: a common practice with boundaries
The technique at the center of the allegations is known as distillation — a process in which the outputs of a more powerful model are fed into a weaker model, training the latter to imitate the stronger one's behavior. Training a frontier-scale model from scratch requires enormous computational resources and investment, which makes large-scale distillation an attractive shortcut for competitors seeking comparable capabilities at a fraction of the cost.
At small scale and conducted openly, distillation is both legal and widely practiced across the AI industry. The objection raised by Kratsios centers on the alleged covert and industrial-scale nature of Moonshot's activity, which he characterized as theft rather than routine engineering.
Wednesday's accusations build on findings Anthropic published in a February report. The company said it traced more than 3.4 million Claude conversations back to Moonshot, routed through hundreds of fabricated accounts. According to Anthropic, some of the account records matched the public profiles of Moonshot's own senior staff. The company said the campaign specifically targeted Claude's capabilities in reasoning, coding, and vision.
Kimi K3 set for full public release July 27
Moonshot released Kimi K3 on July 16. At 2.8 trillion parameters, it is the largest open-weight model released to date. Demand was so high that the company suspended new Kimi sign-ups within two days of launch.
The full model weights are scheduled to go public by July 27. Once released as open weights, they will allow anyone to download, run, and modify the model locally without ongoing oversight from Moonshot or any other party. Anthropic has flagged this release as a risk, arguing that models built through copying may shed the safety guardrails present in the original, and that once weights become openly available, they can spread beyond any party's control.
Broader U.S.–China AI tensions
The dispute adds to a pattern of similar allegations. In January 2025, AI Czar David Sacks said DeepSeek had copied OpenAI's models — a claim DeepSeek rejected outright. Beijing also dismissed Anthropic's February findings as groundless. Moonshot has not commented publicly on the accusations, and Kratsios did not name any specific penalties.
The allegations come as both governments prepare for AI-focused talks scheduled for September. The United States continues to outspend China on AI by nearly 23 to one, though Chinese laboratories have steadily narrowed the performance gap between U.S. and Chinese AI models.
