US officials’ claim that Moonshot AI built its Kimi K3 model by distilling Anthropic’s Fable has triggered a sharp response from the company and growing skepticism among AI researchers, turning a technical dispute into a wider debate over how governments police AI training practices.
Michael Kratsios, director of the White House Office of Science and Technology Policy, has alleged that Moonshot used a “large-scale, covert industrial distillation” platform to tap US frontier models, including Anthropic’s Fable, and then used those outputs to train Kimi K3. He argues this amounted to cloning US capabilities and undermining American research.
Moonshot has not issued a formal, detailed statement on the training data behind Kimi K3, but an employee publicly pushed back on social media by highlighting the narrow window between Anthropic’s Fable launch and K3’s release. The employee noted that Fable went public on July 1 while K3 launched on July 15, and described the idea that Moonshot trained “a brand new frontier model in just 15 days” as implausible, implicitly questioning the US timeline for alleged distillation.
Beyond Moonshot, global AI experts have challenged the core narrative that Kimi K3’s performance can be explained primarily by distillation from Fable. Researchers quoted in multiple reports say there is no publicly presented technical evidence that K3 is a distilled copy, and argue that distillation alone is unlikely to yield such a strong model so quickly after Fable’s release. Some add that as Chinese systems move closer to the frontier and lean more on reinforcement learning, basic supervised distillation becomes a less decisive factor in model quality.
Critics also question the framing of distillation as “theft.” Several AI commentators point out that training smaller models on the outputs of larger systems is a widely used technique in the industry, and that model outputs themselves are typically not treated as copyrighted works. From this perspective, the concern is less about the technical method and more about whether the US is using IP and security rhetoric to justify new restrictions on foreign AI competitors.
Supporters of the White House position, including Anthropic policy executives, maintain that what they call “illicit, adversarial distillation” should be treated as industrial espionage that can strengthen foreign military and intelligence capabilities. They argue that public release of powerful distilled models compounds the harm by making US-derived capabilities globally accessible at low cost.
With Moonshot’s Kimi K3 now seen as one of the most capable open-weight models, the dispute is feeding a broader policy discussion over how far governments can and should go in regulating where training data comes from, how model outputs are reused, and whether open-weight frontier models built outside the US will face sanctions or access limits. For now, the US allegations stand without detailed evidence, while Moonshot and independent experts stress that K3’s competitiveness cannot be reduced to simple claims of copying Fable.
Read Article: Trump Imposes New Tariffs on India and 59 Nations Over Forced Labor Concerns
