Experts challenge US claims that Kimi K3 was built via Anthropic distillation

White House science advisor Michael Kratsios and Treasury Secretary Scott Bessent have levelled serious allegations against Chinese artificial intelligence firm Moonshot, claiming the company developed its Kimi K3 large language model through the illicit distillation of Anthropic’s Fable model and the use of banned Nvidia chips. Kratsios described the alleged activity as "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology," while Bessent stated that US officials were finding "watermarks" of American models within Chinese outputs.

Moonshot has not responded to requests for comment regarding its training process or the specific allegations. Kratsios did not provide further details on the sources of his claims, and the Treasury Department declined to specify what the alleged watermarks consist of. The accusations come amid broader discussions in the US about potentially banning Chinese open-weight models, which have caused significant disruption in the global AI sector.

Despite the high-profile nature of the claims, AI experts are expressing strong scepticism regarding the technical feasibility of distilling a model of Kimi K3’s capability in such a short timeframe. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, told TechCrunch that he did not believe a model of that strength could be produced so quickly following Fable’s release through distillation alone.

Hancock pointed out that Anthropic’s Fable model only became publicly available on 1 July 2026. He argued that it is technically improbable to distill sufficient data, train a new model, and release it within the two weeks that have elapsed since its public debut. The rapid timeline suggests that other factors, beyond simple data copying, likely contributed to Kimi K3’s development.

Nathan Lambert, an AI researcher at the Allen Institute for AI, added that the industry is shifting away from relying solely on distillation or supervised fine-tuning as models approach frontier capabilities. He noted that distilling Fable-like capabilities would likely require reinforcement learning, a process that involves using tens of millions of agents to grade responses. This method is described as "insanely expensive" and potentially a bottleneck due to the slow speed of frontier models.

The dispute also highlights the difficulty in separating allegations of data theft from claims that Moonshot accessed banned Nvidia Grace Blackwell 300 chips and servers in Thailand. Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology, noted that while a black market for advanced chips exists, disentangling the distillation claims from the hardware smuggling allegations remains complex. The founder of server builder Supermicro was indicted in May for allegedly smuggling advanced chips into China.

While distillation is a common practice in the industry, with Elon Musk testifying that his company SpaceXAI distilled OpenAI models to develop Grok, experts argue that American officials may be underestimating the technical expertise of Chinese teams. Hancock emphasised that Moonshot’s founders are legitimate researchers and engineers doing solid work, suggesting that China’s progress would continue even if American models were to stall.