Allegations of Covert Model Distillation

The White House Office of Science and Technology Policy Director Michael Kratsios stated that US officials possess information indicating Moonshot AI distilled Anthropic’s Fable model during the development of its Kimi K3 system. According to Kratsios, the Chinese firm constructed a dedicated internal platform designed for large-scale extraction of capabilities from American models.

This platform reportedly enabled Moonshot to rotate between different access methods rapidly, aiming to evade detection. Kratsios emphasized that while model distillation represents a standard technique for creating smaller, more efficient systems, the alleged industrial-scale and secretive application crosses into unacceptable territory.

Kratsios further noted that Moonshot increased its computing resources by acquiring NVIDIA GB300 servers, positioned in Thailand, for AI training purposes. These advanced chips are subject to US export restrictions preventing direct transfer to China.

US Policy Response and Potential Consequences

US Treasury Secretary Scott Bessent reinforced the administration’s stance, distinguishing between support for open-source AI innovation and what he described as exploitation of American intellectual property. He indicated that covert operations crossing into IP theft could trigger sanctions and additions to the Entity List.

The accusations arrive amid broader US efforts to balance AI advancement with national security concerns, including recent adjustments to export controls on high-performance models.

Kimi K3’s Rapid Rise and Technical Context

Moonshot AI launched Kimi K3 on July 16, positioning it as a 2.8 trillion-parameter open-weight model focused on advanced reasoning, coding, and knowledge tasks. The release generated significant demand, prompting the company to temporarily halt new user sign-ups. Full model weights are scheduled for public release on July 27.

The timing has drawn attention, as Anthropic’s Fable 5 became available again on July 1 following a brief period offline due to export control considerations. Some independent researchers have questioned whether distillation from Fable could fully account for Kimi K3’s reported performance levels within such a short window.

Elie Bakouch of Prime Intellect and Dean Ball of OpenAI have publicly expressed skepticism about the technical feasibility of distillation explaining the model’s capabilities in this timeframe.

Prior Concerns from Anthropic

The latest claims build on earlier reports from Anthropic, which in February documented millions of interactions with its Claude models traced back to Moonshot AI through fabricated accounts. Those efforts reportedly targeted strengths in reasoning, coding, and vision capabilities.

Moonshot AI has not issued a public response to the White House allegations. No supporting evidence has been released publicly by US officials to allow independent verification.

Broader Implications for US-China AI Competition

The dispute highlights ongoing friction between the United States and China in artificial intelligence development. While the US maintains a substantial lead in overall AI investment, Chinese laboratories continue to narrow performance gaps with frontier models.

Industry observers note that such accusations could influence future regulatory approaches, enforcement priorities, and international collaboration on AI technologies. The episode underscores the challenges of protecting proprietary capabilities in an era of rapid model iteration and distributed computing resources.