Kimi K3 Distillation: What the White House Claims

Kimi K3 distillation claim: what Michael Kratsios posted

On July 22, 2026, Michael Kratsios said on X that the US has "information" that Moonshot AI distilled Fable to build Kimi K3. Kratsios directs the White House Office of Science and Technology Policy (OSTP). Fable is Anthropic's model, sold as Claude Fable 5.

The post at a glance

Posted
July 22, 2026, on X, from the account @mkratsios47
Author
Michael Kratsios, Director, White House Office of Science and Technology Policy
Subject
The Chinese AI company Moonshot AI and its model Kimi K3
Claim 1
There is information that Anthropic's Fable was distilled to develop K3
Claim 2
An internal platform was built to switch access methods and avoid detection
Claim 3
GB300-equipped servers were acquired, and GB300s accessed in Thailand

Moonshot AI distillation of Fable, in the post's own words

The first paragraph of the post makes the core claim: Moonshot AI distilled Anthropic's Fable for the development of K3. It is a short post on X, not a released report or a legal filing. It gives no time span and no volume for the activity.

The second paragraph gives the method. Moonshot is said to have built an internal platform for large-scale distillation against US models, one that switches quickly between methods of access to avoid detection.

We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model. — The opening paragraph, which carries the claim
View official source →

GB300 servers in Thailand: the chip claim

The chip claim: Moonshot acquired GB300-equipped servers and accessed GB300s in Thailand. The hedge "likely" is the post's own and covers only the purpose: the post says the GB300s were likely used to train Moonshot's AI models.

The post says "GB300-equipped servers" and goes no further. It does not name a chipmaker or Alibaba, does not say how the servers were obtained, and does not explain why Thailand comes into it. Anything beyond those words is interpretation, not part of the post.

To do this they 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. Moonshot AI has also acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models. — The method and the chip claim
View official source →

AI distillation versus a distillation attack: the US line

The post has no quarrel with distillation as a method. Its target is the large-scale, covert, industrial kind meant to steal US technology.

Anthropic defines distillation as training a less capable model on the outputs of a stronger one. It calls the method widely used and legitimate, since frontier labs routinely shrink their own models this way into cheaper versions. For a fuller account, see AI Distillation Attacks in PLA-Linked Research.

Legitimate use versus an adversarial distillation attack

In the post's words, legitimate distillation "plays a vital role," while covert industrial distillation is unacceptable. The US, it says, backs a competitive ecosystem of open-weight models, open-source frameworks, specialized systems and frontier models. Open-weight means the trained parameters are published for download.

Anthropic uses the name distillation attack for the illicit kind. The post's test has three parts: scale, concealment, and an aim of stealing proprietary technology.

The United States strongly supports the free and fair development of AI, including a thriving competitive ecosystem that spans frontier models, specialized systems, open-source frameworks, and open-weight models. Legitimate AI distillation used to create smaller, more efficient models plays a vital role in this open innovation ecosystem. However, large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable. — The passage that separates legitimate distillation from the kind treated as a problem
View official source →

Anthropic distillation report: 3.4 million exchanges

Anthropic's February report names Moonshot, but it never mentions K3 or Fable, so it does not back the White House claim. On February 23, 2026, Anthropic said DeepSeek, Moonshot and MiniMax had run industrial-scale campaigns against Claude. It counted over 16 million exchanges through about 24,000 fraudulent accounts.

Moonshot's share was over 3.4 million exchanges through hundreds of fraudulent accounts. The targets were agentic reasoning and tool use, coding and data analysis, computer-use agent development and computer vision. Anthropic attributed the campaign through request metadata that matched the public profiles of senior Moonshot staff.

The February report predates both K3 and Fable's July availability. It is a separate finding about earlier activity against Claude, and it neither confirms nor contradicts the July post.

Moonshot (Kimi models) employed hundreds of fraudulent accounts spanning multiple access pathways. Varied account types made the campaign harder to detect as a coordinated operation. We attributed the campaign through request metadata, which matched the public profiles of senior Moonshot staff. In a later phase, Moonshot used a more targeted approach, attempting to extract and reconstruct Claude’s reasoning traces. — The passage on Moonshot
View official source →

Is Kimi K3 distilled from Claude? The evidence so far

As of July 31, 2026, the public record holds a claim and the reactions to it, but no published evidence for it. Claude Fable 5 has been available worldwide again since July 1, 2026; the background is in Is Claude Fable 5 Back? Yes, With New Safeguards.

No evidence in the White House post: what is missing

The post offers no evidence: it says "we have information" and stops. Missing evidence is a fact about the post, not a finding on whether the claim is true.

Not in the post

Evidence
No supporting document and no method of confirmation
Measures
No legal action or sanction announced in the post
The other side
No quote from Moonshot AI, Anthropic or any chipmaker
Scale
"Large scale" is asserted, with no time span or volume
Thailand
No explanation of why Thailand or how the servers were obtained

CyberScoop reported that Kratsios gave no details on how the government obtained its information. It also said a request for comment sent to Moonshot was not returned before publication.

Kratsios did not provide details on how the U.S. government learned that K3 had been distilled from Anthropic’s model. — CyberScoop on the basis of the claim
View official source →

Moonshot's denial and the dispute over the timeline

Moonshot denies using American models to train Kimi K3, and analysts question whether the timeline allows it. CNN reported that Huang Zhenxin, Moonshot's head of enterprise business, denied the accusations in an interview with state-run media on the Monday before its July 23 report. That was July 20, two days before the post, and CNN does not say the denial answered it.

A Moonshot employee, Randy Xian, wrote on X that Fable was released on July 1 and K3 on July 15, which would mean training a new frontier model in "just 15 days," per SCMP. That is his claim, not a date Moonshot has confirmed.

Wei Sun, an analyst at Counterpoint Research, told CNN that Moonshot would not have had enough time to train on its rivals' latest models, and called the timeline unconvincing. CNN's own summary is that most analysts found the accusation unlikely.

SCMP reported that AI experts posting on X on July 23 called the accusations political and reckless, and noted that a model's output is not copyrighted. If Moonshot publishes a fuller statement, put its wording side by side with the post to see what changed between the two accounts.

Anthropic, Treasury and the sanctions warning

Sarah Heck, Anthropic's head of policy, replied to Kratsios on X and accused Moonshot of "IP theft and industrial espionage," per CNN. Treasury has also warned of sanctions, though without naming a company.

Treasury Secretary Scott Bessent said on X on July 22 that sanctions and Entity List designations are possible if Chinese firms run secret, industrial-scale distillation that amounts to IP theft. CNN notes that he named no firm, Moonshot included. A warning that sanctions are possible is not a sanction.

Kimi K3 Alibaba Nvidia chips: what Bloomberg reported

The Bloomberg report is a separate story from the GB300 claim, and it rests on anonymous sources. Mackenzie Hawkins and Haze Fan published it in Bloomberg on July 31, 2026.

Moonshot AI Alibaba: a computing-power agreement

Bloomberg reported that Moonshot has a computing power deal with Alibaba covering around 20,000 Nvidia chips. The sources are people with knowledge of the companies' operations, who asked for anonymity because the deal is confidential. They said the hardware is a substantial share of the computing capacity Moonshot uses for its Kimi models.

Bloomberg frames the deal as a sign of China's continued reliance on Western semiconductors, and that framing is the outlet's own. The report's opening names no chip model and no supply route.

Chinese AI champion Moonshot has a computing power agreement with Alibaba Group Holding Ltd. for the use of around 20,000 Nvidia Corp. chips, underscoring China’s continued reliance on Western semiconductors to fuel its AI development, according to people with knowledge of the companies’ operations. — The opening paragraph of the report
View official source →

Two chip stories from two different sources

The White House post and the Bloomberg report are two separate claims, and merging them into one storyline goes past what either one says. They differ in source, date, hardware and partner, as “Two chip stories side by side” shows.

Two chip stories side by side

White House post Bloomberg report
Source Kratsios's post on X Anonymous people with knowledge of the operations
Date July 22, 2026 July 31, 2026
Hardware named GB300-equipped servers Around 20,000 Nvidia chips
Place or partner Thailand Alibaba
Mentions the other Neither Alibaba nor Nvidia Not in the readable text: GB300, Thailand or the White House post

Anthropic's February report does connect distillation and chips in general terms. It argues that extraction at scale needs advanced chips, but that is an argument about export controls, not a finding about Moonshot's servers. For Kimi K3 itself, see Kimi K3 benchmarks, pricing and open weights.

What the record supports about Moonshot so far

As of July 31, 2026, the public record shows a White House claim with no published evidence, and nothing yet that settles it. Four points hold up.

  • The post is a claim: it shows no evidence, no method of confirmation and no legal action.
  • Moonshot denies using American models to train K3, and analysts question the timeline.
  • Anthropic's February report counts over 3.4 million Moonshot exchanges, but not K3 or Fable.
  • The Bloomberg compute report comes from anonymous sources and stands apart from the GB300 claim.

In the Kimi K3 distillation dispute, the next thing to watch is whether the US shows its evidence or Moonshot answers the post directly. Until then, the post stays a claim rather than a finding.

FAQ

What exactly did the White House claim about Moonshot AI?
Michael Kratsios, who directs the White House Office of Science and Technology Policy, posted on X that the US has information that Moonshot AI distilled Anthropic's Fable to develop its K3 model. He added that Moonshot built an internal platform that switches quickly between methods of access to avoid detection, and that it acquired GB300-equipped servers and accessed GB300s in Thailand. The post hedges only the purpose of the GB300 use, saying it was "likely" training.
Is the US objecting to distillation itself?
No. The post says legitimate distillation, used to create smaller and more efficient models, plays a vital role in the open innovation ecosystem. What it calls unacceptable is large-scale, covert industrial distillation aimed at stealing proprietary US technology and undermining American research. The line it draws is about scale, concealment and intent.
Has anyone shown evidence that Kimi K3 was distilled from Fable?
Not in the public record as of July 31, 2026. The post says only that the administration has information, and CyberScoop reported that Kratsios gave no details on how the government learned it. Moonshot has denied using American models to train K3, and Anthropic's February report on Moonshot does not mention K3 or Fable.