The White House Says Kimi K3 Was Distilled From Claude. The Same Day, 200 Startups Begged Washington Not to Ban Chinese Open Weights.

TL;DR

The White House just named a specific Chinese lab, a specific American model, and a specific offense. Michael Kratsios, director of the Office of Science and Technology Policy, accused Moonshot AI on Wednesday of building Kimi K3 through "industrial distillation" of Anthropic's Claude Fable 5, and of training on Nvidia GB300 servers that export controls say Chinese firms cannot have. Treasury Secretary Scott Bessent followed with a warning that sanctions and Entity List designations are "on the table." The same afternoon, per Politico, nearly 200 startups sent letters begging the administration not to cut off Chinese open-weight models. K3's weights are still promised for July 27. This is now the biggest open question in open weights.

The accusation

Kratsios posted the allegation on X: Moonshot, he wrote, "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." In a statement to CyberScoop he called it "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research."

The nuance matters, and to his credit Kratsios drew it: distillation itself is a normal, legal training technique. You feed a stronger model's outputs to a weaker model and train the weaker one to imitate it, the way every lab shrinks its own frontier models into cheap fast ones. The accusation is not "Moonshot used distillation." It is that Moonshot ran a covert harvesting operation against a competitor's API at industrial scale, which is less like a student learning from a textbook and more like enrolling 24,000 fake students in a rival's masterclass and pooling their lecture notes into a curriculum.

Moonshot did not respond to requests for comment from TechCrunch or CyberScoop. Neither Kratsios nor Treasury published the underlying evidence for the Fable-specific claim.

The paper trail that already existed

The reason this accusation lands harder than a normal Washington broadside: Anthropic published a detailed forensic report on February 23 titled "Detecting and preventing distillation attacks." It named three labs, DeepSeek, Moonshot, and MiniMax, and put numbers on the harvesting: over 16 million exchanges with Claude through approximately 24,000 fraudulent accounts, "in violation of our terms of service and regional access restrictions."

Per that report, Moonshot's 3.4 million-plus exchanges targeted agentic reasoning and tool use, coding and data analysis, computer-use agent development, and computer vision. Which is, not coincidentally, a decent summary of what K3 is good at. The report's best detail: Anthropic tied the fraudulent accounts to Moonshot partly through "request metadata, which matched the public profiles of senior Moonshot staff." The opsec equivalent of robbing a bank with your employee badge clipped to the ski mask.

The chips

Kratsios stacked a second allegation on top: Moonshot "acquired Nvidia's GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models." GB300 is Blackwell-generation hardware that is banned from sale to Chinese companies under US export controls, so if that claim holds, it is a separate violation with its own enforcement machinery, no distillation debate required.

Bessent's statement to TechCrunch tied it all together: "Open source is not open season on American IP. When firms conduct covert, industrial-scale distillation attacks that cross the line into IP theft, sanctions and Entity List designations will be on the table." Representatives Andrew Garbarino and John Moolenaar announced a joint congressional investigation into what they called "large-scale theft of proprietary capabilities."

The timeline problem

Now the part the press releases skip. Claude Fable 5 has been publicly available since July 1. Kimi K3, a 2.8-trillion-parameter model, launched July 16. If your mental model of the claim is "Moonshot distilled Fable into K3," fifteen days is not enough time to pretrain a frontier-scale model from anyone's outputs, and skeptics pointed that out within hours of the accusation.

The more defensible reading, and the one Anthropic's own February data supports, is that the harvesting is a long-running program against Claude models generally, months before Fable existed, and that distilled outputs feed post-training rather than pretraining. Post-training is where a model learns to be an agent: tool-calling traces, reasoning chains, coding sessions. Two weeks of harvesting a new frontier model can absolutely move the needle there. But "K3's polish came partly from distilled Claude data" is a much narrower claim than "K3 is stolen Fable," and the administration's evidence for either version is, so far, not public.

The counter-lobby arrived the same day

While Kratsios was posting, the newly formed Little Tech Association delivered letters to President Trump, Commerce Secretary Howard Lutnick, Secretary of State Marco Rubio, Bessent, and Kratsios himself, signed by nearly 200 companies including Y Combinator and Proton. Their position, per the letter Politico obtained: "American leadership requires two things: world-leading American open-weight models and continued access for U.S. builders to open models already available worldwide."

The trigger is real: Axios reported Monday that the administration has been fighting internally over restricting Chinese models since K3 launched. The association's executive director, Harry Godfrey, asked policymakers to use "a scalpel rather than a sledgehammer." Startup founder Suhail Doshi was blunter: "There'll be hundreds of companies that instantly die" if the models are banned.

What this means for you

If you build on Chinese open-weight models, and given Qwen, DeepSeek, and Kimi's market share, you statistically do, this is the week the bill for cheap frontier-adjacent weights arrived in Washington. Three things to watch:

• July 27. Moonshot has promised K3's checkpoints on its Hugging Face org under a reported modified-MIT license. Shipping open weights into the teeth of a sanctions threat would be a statement. Delaying would be a bigger one, especially with a $30 billion Hong Kong IPO riding on the open-weight narrative.
• The enforcement gap. An Entity List designation blocks business relationships with US firms, which could hit distribution channels, cloud hosts, and partnerships. What it cannot do is un-download a checkpoint. Washington is threatening a company; the artifact, once released, answers to no one.
• The precedent. If "covert industrial distillation" becomes a sanctionable category, every frontier lab's terms of service just acquired a foreign-policy enforcement arm. That cuts both directions, and the startups signing Wednesday's letter clearly noticed.

Key Takeaways

• OSTP director Michael Kratsios publicly accused Moonshot AI of building Kimi K3 via industrial-scale distillation of Anthropic's Fable and of training on export-banned Nvidia GB300 servers, including via access in Thailand.
• Treasury Secretary Bessent says sanctions and Entity List designations are "on the table"; Reps. Garbarino and Moolenaar opened a joint congressional investigation.
• Anthropic's February report documented 16M+ Claude exchanges across ~24,000 fraudulent accounts, with 3.4M+ traced to Moonshot, partly via metadata matching senior staff profiles.
• The timeline undercuts the strongest version of the claim: Fable was public for only 15 days before K3 launched, so distillation could plausibly shape post-training, not pretraining.
• Nearly 200 startups, including Y Combinator and Proton, sent letters the same day urging the administration not to ban Chinese open-weight models.
• K3's open weights are still promised for July 27; whether they ship on schedule is now a geopolitical signal, not just a release date.

Sources: TechCrunch, CyberScoop, Anthropic: Detecting and preventing distillation attacks, Politico, Axios