On July 22, White House Office of Science and Technology Policy Director Michael Kratsios posted a fairly sharp statement on X: the US has information that Chinese startup Moonshot AI covertly copied Anthropic's Fable model to build its new Kimi K3. Within hours, Treasury Secretary Scott Bessent picked up the thread and threatened sanctions. The story broke fast and loud - and, as it turns out, on a fairly shaky evidentiary basis.
What distillation is, and why it's usually legal
The technique in question is called model distillation. The idea is simple: take a powerful model, feed it queries, collect its responses, then train a weaker model on those responses so it learns to imitate the stronger model's style and quality. This is a standard, widely accepted technique in the industry when done openly and at small scale - it's how practically every major lab optimizes and cheapens its models.
The problem arises when it's done covertly, at industrial scale, bypassing a competitor's API access restrictions. That's exactly what Kratsios accused Moonshot of: according to him, the company built a dedicated internal platform for large-scale distillation against US models, capable of quickly switching between multiple access methods to avoid detection. He separately added an allegation that Moonshot had gained access to restricted Nvidia GB300 chips (Blackwell generation, banned from sale to China) - reportedly through Thailand.
Treasury's response, and the word "sanctions"
Bessent posted a line the same day that's already being widely quoted: "Open source is not open season on American IP." He added that covert, industrial-scale distillation attacks could lead to sanctions and Entity List designations - effectively a blacklist that cuts off access to American technology and suppliers.
Meanwhile, Representative Ted Lieu immediately pointed out a contradiction in the administration's position: the government is simultaneously threatening sanctions over stolen technology while still allowing sales of powerful AI chips to China.
The accusation's biggest problem - the timeline doesn't add up
Here's where the story gets genuinely interesting. Kratsios provided no technical evidence or details on exactly how the government learned about the distillation - just the statement itself, posted on social media.
Experts immediately flagged a timing mismatch. Fable only became publicly available on July 1, 2026. Kimi K3 shipped on July 16 - meaning, if the accusation is accurate, Moonshot had exactly two weeks to carry out what Kratsios called "large-scale industrial distillation against US models." Technically, that's a remarkably tight window for a process that usually requires far more time and computational resources. Some researchers are openly questioning whether K3 could have been built primarily through distilling Fable in such a short span at all.
There's also broader context worth noting: Anthropic itself had separately stated back in February that it traced 3.4 million Claude exchanges to the same startup - meaning suspicions predated the K3 release by months, even though that doesn't prove this specific accusation.
What Kimi K3 actually is
The irony of the situation is that the model at the center of the scandal is genuinely strong. It's 2.8 trillion parameters, the largest open model to date, with a 1-million-token context window and built-in vision support. Independent evaluations place it right behind Fable 5 and GPT-5.6 Sol among the strongest available systems - and precisely because it's open-weight rather than closed, the model's full weights become public on July 27.
That, incidentally, is part of the substance of the US complaint: if a model was built on a stolen foundation and then released for free, the argument goes, the harm to the original developer only compounds, since anyone can now download the copy.
What I take away from this
The situation looks like a textbook case of geopolitics moving faster than technical evidence. The accusation landed loudly, with specific-sounding details (an internal platform, rotating access methods, chips routed through Thailand), but without any public confirmation of those details - just an official's statement on social media. Meanwhile, the simplest counterargument, a plain two-week timeline, calls into question whether the accusation is even technically plausible as stated.
That doesn't mean the accusation is necessarily false - Anthropic had its own suspicions well before this moment, and the full picture may be more complicated than what's publicly visible right now. But as long as the official version rests on an unsupported statement, and the most obvious counterargument, basic date arithmetic, goes unanswered, it seems reasonable to wait for something more concrete than another post on X before drawing conclusions.