White House science and technology chief Michael Kratsios has accused China’s Moonshot AI of using large-scale distillation to train its Kimi K3 model, saying the lab siphoned outputs from Anthropic’s Fable 5. The allegation comes just as Moonshot prepares to open-source the weights of K3, a frontier-tier model it released last week. Kratsios framed the practice as industrial-scale extraction of proprietary US technology rather than legitimate model compression.<\/p>\n
According to the source report, Kratsios claimed that Moonshot built a sophisticated internal platform to run large-scale distillation against US models. To avoid detection, the platform allegedly switched between multiple methods of access while pulling from frontier US systems.<\/p>\n
The same report states that Moonshot acquired GB300-equipped Nvidia servers and has accessed GB300 hardware in Thailand, likely to train its AI models.<\/p>\n
Kratsios distinguished legitimate distillation, which produces smaller and more efficient models within an open innovation ecosystem, from covert industrial distillation aimed at stealing proprietary US technology. His full statement said: “The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale, covert industrial distillation aimed at stealing proprietary US technology and undermining American research is unacceptable.”<\/p>\n
Sarah Heck of Anthropic’s policy team described distillation as “IP theft and industrial espionage” and tied the issue to national security risks. Anthropic had previously accused Moonshot, along with DeepSeek and MiniMax, of “industrial-scale campaigns” to “illicitly extract Claude’s capabilities” in February.<\/p>\n
Moonshot’s Kimi K3 was released last week with benchmark scores at or near the US frontier. The full model is scheduled to have its weights published on July 27. The open release is what makes the timing of the distillation allegation politically charged: a model positioned as an open-weight Chinese frontier release is now being framed by the White House as a product of illicit extraction.<\/p>\n
The source report notes that concrete proof of distillation has not been disclosed, and that the accusation also fits earlier predictions that the US would move against open Chinese models. If the claim holds up, the lens on K3’s release shifts from a meaningful open-weight step forward to a potentially illegally siphoned geopolitical flashpoint. The Thailand-routed GB300 access adds a hardware dimension, suggesting that the alleged distillation campaign required frontier-class compute that Moonshot worked to obtain outside standard channels.<\/p>\n
White House science and technology chief Michael Kratsios accused Moonshot AI of distilling Anthropic’s Fable 5 to train Kimi K3.<\/p>\n
He said Moonshot built an internal platform to run large-scale distillation while rotating access methods to avoid detection, and that Moonshot obtained GB300-equipped Nvidia servers and accessed GB300s in Thailand, likely for AI training.<\/p>\n
Kimi K3 was released last week, with the full model’s weights scheduled to be published on July 27.<\/p>\n