The White House Says Kimi K3 Is Distilled Fable. The Evidence Lives in Anthropic's Server Logs.

TL;DR

On 22 July, OSTP director Michael Kratsios accused Moonshot of running a covert platform to distill Anthropic’s Fable and build Kimi K3, days before K3 goes open-weight and while Treasury floats sanctions. Distillation is a real, legitimate technique, and doing it against a rival is cheap: Berkeley’s Sky-T1 matched o1-preview reasoning for under $450 and 19 hours of fine-tuning. So the story is plausible as a capability boost, since fine-tuning on captured reasoning traces is fast and the base model does the heavy lifting. But nobody can prove distillation from the weights alone. The actual evidence, from Anthropic’s February report, is 3.4 million exchanges through fraudulent accounts, all traffic-side. Kratsios added an accusation, not a forensic case.

Michael Kratsios runs the White House Office of Science and Technology Policy, and on Wednesday morning he posted something a cabinet-level science adviser does not usually post: a specific allegation of industrial espionage against a named foreign company. “We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model,” he wrote, adding that Moonshot built “a sophisticated internal platform to conduct large scale distillation against U.S. models” and had accessed Nvidia GB300s in Thailand to train on. The post cleared 14,000 likes. The top reply, from a developer named Haider, read: “we have information that Anthropic distilled humanity’s entire internet for the development of Fable 5.”

The timing is the first thing worth clocking. Kimi K3 releases its full weights next week, and it already benchmarks at or above the leading US frontier models. Treasury Secretary Scott Bessent had just suggested sanctions on Chinese labs if improper distillation could be proven. So the accusation lands exactly when it is most useful and least testable.

What “distilling a model” actually means

Distillation, in the textbook sense, is training a small student model to imitate a large teacher. Geoffrey Hinton’s 2015 formulation had the student match the teacher’s full probability distribution over tokens, the “soft labels” that carry more information than a single right answer. Labs do this constantly and openly: every cheap, fast model in a product lineup is usually a distilled version of a bigger sibling.

Here is the catch that the word “distillation” papers over. You cannot do Hinton-style distillation against a competitor’s API, because an API hands you generated text, not the underlying probabilities. What you can do is collect the teacher’s outputs at scale and fine-tune your own model on them. That is sequence-level distillation, or, more honestly, synthetic-data training. It transfers specific capabilities (reasoning chains, tool-use patterns, coding style) onto a base model you already trained yourself. It does not clone Fable. It borrows Fable’s homework.

That distinction matters for the accusation. “Moonshot distilled Fable for K3” most plausibly means Moonshot used captured Fable reasoning traces to sharpen K3’s agentic and coding skills, not that K3 is a copy of Fable wearing a disguise. Anthropic’s own February writeup said as much: in a later phase, Moonshot moved to “extract and reconstruct Claude’s reasoning traces,” which is exactly the recipe for teaching a model to think out loud.

Is the timeline even plausible?

Yes, and that is the uncomfortable part. The expensive, slow work in building a model is pretraining the base. Fine-tuning on distilled traces is neither. In January 2025 a Berkeley team trained Sky-T1, a 32B reasoning model that matched OpenAI’s o1-preview on several benchmarks, by fine-tuning an existing open model on distilled reasoning data. Cost: under $450. Time: 19 hours on eight H100s. That is the whole point of distillation, and it is why export hawks lose sleep over it: the capability transfer is orders of magnitude cheaper than the capability.

So the bottleneck for Moonshot was never the fine-tuning run. It was capturing enough high-quality Fable output without getting cut off. Anthropic says Moonshot generated over 3.4 million exchanges through hundreds of fraudulent accounts across multiple access pathways, targeting agentic reasoning, tool use, coding, and computer vision. When Anthropic caught MiniMax running the same playbook and shipped a new model mid-campaign, MiniMax redirected nearly half its traffic to the new model within 24 hours. These operations are industrialized. A months-long capture window followed by a fast fine-tune is entirely consistent with K3 arriving when it did.

The proof problem

None of that is proof that K3 specifically is distilled Fable. Here is the thing about distillation: you cannot demonstrate it from the student’s weights. A model trained on another model’s outputs looks, from the outside, like a model that trained well. There is no watermark in a reasoning trace. The only durable evidence sits on the teacher’s side, in the API logs, the account metadata, the payment patterns. That is precisely where Anthropic’s case lives, and it is strong: IP correlation, request metadata matching the public profiles of senior Moonshot staff, coordinated timing across accounts.

Kratsios has none of that in public. He has “information.” His statement is an intelligence assertion dressed as a technical finding, and it inherits the shape of the DeepSeek episode from January 2025, when OpenAI told the FT it had “seen some evidence” of distillation and White House adviser David Sacks claimed “substantial evidence” on Fox News without producing any. Same accusation, same absence of a published forensic trail, same convenient news cycle.

The politics under the technology

The other tell is who is nodding along. The Trump administration and Anthropic have spent 2026 at odds, with export controls that blocked Fable 5’s release and a Pentagon move to blacklist the company in February. On distillation, that fight evaporates. Anthropic’s head of public policy thanked Kratsios within the hour and called adversarial distillation “IP theft and industrial espionage.” A rival’s cheap model becomes a national-security problem, and everyone with a frontier API to sell agrees at once.

The real risk Anthropic named is worth keeping: a distilled model inherits capabilities without inheriting safeguards, and an open-weight one spreads them with no off switch. That concern is legitimate, and it connects directly to what happened when OpenAI’s own eval model went off the rails with its safeguards deliberately switched off. But “distilled models are dangerous” and “K3 is provably stolen Fable” are separate claims. The first has evidence. The second, so far, has a tweet.

Sources

Related on this blog