Kratsios says Moonshot built Kimi K3 through 'industrial distillation' of Anthropic's Fable

The China-based startup Moonshot AI developed its latest mode, Kimi K3, through the "industrial distillation" of Anthropic's (ANTHRO) Fable,  according to Michael Kratsios, the science advisor to U.S. President Donald Trump.

"We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," Kratsios said in a post on X Wednesday morning.

What's more, Moonshot also gained access to Nvidia's (NVDA) GB300 through servers located in Thailand to help train its models. Under U.S. Department of Commerce rules, Nvidia's GB300 chips are not allowed for export to China. 

"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 access to avoid detection," Kratsios said. "Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models."

On Tuesday, U.S. Treasury Secretary Scott Bessent said the Trump Administration is entitled to sanction open-weight models from China if evidence of distillation from American models is verified.

"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," Kratsios added. "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 U.S. technology and undermining American research is unacceptable."

Distillation is a training method where less capable models are improved by utilizing outputs from stronger models.

Kimi K3 caused a stir through the AI market when it was released last week. Moonshot said the 2.8 trillion-parameter system is the world's largest open-weight AI model. Kimi K3 "performed competitively" with Fable 5 and "substantially outperformed" OpenAI's (OPENAI) Opus 4.8, GPT 5.6 Sol and GPT 5.5 in terms of GPU kernel optimization, the company said.