White House science chief Kratsios accuses Moonshot of distilling Claude Fable to build Kimi K3 and of using GB300 servers in Thailand
Partly confirmed
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 3 of 5
- Last verified
- 10 October 2026
Your AI and this story
- GPT-6 Astra83 days after its cutoff
- Claude Opus 5.522 days after its cutoff
- Gemini 3.8 Flash113 days after its cutoff
- Grok 4.752 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 22 days before it.
Key facts
- Kratsios (X, July 22): ‘We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.’ ‘They built a sophisticated internal platform to conduct large-scale distillation against U.S. models, allowing them to quickly switch between multiple access methods to avoid detection.’ (Forklog)
- Kratsios separated ‘legitimate AI distillation’ from ‘large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research’ (Implicator)
- Chip allegation: Moonshot acquired GB300-equipped servers and accessed GB300s in Thailand, likely to train its models (Implicator)
- No public evidence: no query logs, training-data manifest or forensic analysis were published, and the government did not say how it obtained the information
- Bessent (X, July 22): sanctions or Entity List designations ‘will be on the table’ for covert distillation; a day earlier he had told Fox Business ‘what we do not support is IP theft’
- Responses: Anthropic policy chief Sarah Heck thanked Kratsios and called the conduct ‘IP theft and industrial espionage’; Moonshot, the Chinese Embassy and Nvidia did not comment (Implicator, citing Reuters)
- Sceptics: only ~15 days separated Fable’s public availability (July 1) and K3’s preview (July 16); OpenAI’s Dean Ball said distillation ‘plays a role’ but ‘clearly it’s not primary’
- Later: OpenAI’s Sept 30 threat report described a Moonshot-linked distillation campaign starting July 1 against its own models (entry OpenAI says Moonshot AI-linked individuals ran a coordinated campaign to extract its models’ hidden reasoning)
What happened
Kimi K3, a 2.8-trillion-parameter open-weight model, came out on July 16, 2026 and scored close to US frontier models. Six days later the director of the White House Office of Science and Technology Policy, Michael Kratsios, publicly accused its maker Moonshot AI of building it by “distilling” Anthropic’s Claude Fable, i.e. training on Fable’s outputs at scale, and of using Nvidia GB300 servers that export rules bar from China. The Treasury Secretary threatened sanctions the same day.
The X posts themselves were not read for this entry (X is not fetched); quotes come from Forklog and Implicator. Both say no evidence was made public.
Why it matters
It was the first time a senior US official named a specific Chinese lab and model in a distillation accusation. It set the frame for the later NSA/CISA/FBI advisory on distillation (Sept 8) and OpenAI’s own Moonshot report (Sept 30), and it tied distillation to export-control enforcement and sanctions.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 3 press, 1 reaction
Press
- Implicator: White House accuses Moonshot of distilling Fable and using banned Nvidia chipsimplicator.ai, press
- Forklog (July 23): White House official accuses Moonshot AI of distilling Anthropic’s Fable for Kimi K3forklog.com, press
- The New Stack: Kimi K3, White House alleges Fable 5 siphoningthenewstack.io, press
Reactions
- Pasquale Pillitteri: Kimi K3 distilled from Fable? A US claim without proofpasqualepillitteri.it, discussion
Changes
- Filed (the claim was previously seen only in a search-engine summary)