
Global AI Experts Push Back on US Distillation Claims Against Moonshot's Kimi K3
Researchers questioned the feasibility of distilling Anthropic's Fable into K3 in a 15-day window, as Beijing's AI community rallied around Moonshot and the evidence war moved to timelines and token counts.
A day after the White House accused Moonshot AI of building Kimi K3 by covertly distilling Anthropic's Fable model, a growing chorus of researchers — in China and outside it — is questioning whether the timeline makes technical sense.
The 15-day problem
The South China Morning Post reported that AI experts across the industry are pushing back on the core claim. Anthropic's Fable went public on July 1; Moonshot launched K3 on July 15. Critics argue that a two-week window is implausibly short to run a distillation campaign at the scale alleged, integrate the harvested outputs into a training corpus, and complete a full frontier training run — even before accounting for the months of pretraining a 2.8-trillion-parameter mixture-of-experts model requires.
Moonshot employee Randy Xian made the point sardonically on X: if the accusation were true, he wrote, Moonshot trained "a brand new frontier model in JUST 15 DAYS. Guinness World Record stuff."
What the skeptics concede
The pushback is not a blanket exoneration. Researchers note that distillation does not require the teacher model to be public for the student's entire training cycle — post-training, fine-tuning and preference optimization happen late and could theoretically incorporate recent model outputs. Anthropic's report of some 16 million fake-account interactions originating from Chinese labs also describes conduct that predates Fable's release and spans multiple US models.
What critics dispute is the leap from "Chinese labs harvest US model outputs" — a practice widely assumed to be endemic across the industry, in both directions — to "K3 is a stolen copy of Fable," a far stronger claim that the White House has so far supported with assertions rather than published evidence.
Beijing's counter-narrative
Chinese AI commentators have seized on the ambiguity, framing the accusation as an attempt to discredit an open-weight model that outperformed US rivals on public leaderboards days before its weights ship to the world. With K3's open release still scheduled for July 27, the dispute is no longer academic: enterprises worldwide are about to download the weights, and whether they see a triumph of Chinese engineering or contested intellectual property may depend on evidence neither side has yet put on the table.
Newsletter
Get Lanceum in your inbox
Weekly insights on AI and technology in Asia.


