Kimi K3: White House alleges Fable 5 siphoning
The New Stack Amanda Caswell ● Covered by 75 sources
White House official Michael Kratsios says Moonshot secretly trained its new Kimi K3 model on stolen Anthropic Fable 5 outputs. No public proof yet, but the accusation could trigger tighter chip and API export rules.
Based on reporting by The New Stack, Amanda Caswell — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Michael Kratsios didn't mince words on X this week. The White House technology director claims Moonshot AI built a purpose-made internal platform to pull data out of Anthropic's Fable 5 model, switching between multiple access methods specifically to dodge detection, then used that harvested output to train its new Kimi K3 system. He called it covert industrial distillation aimed at stealing proprietary American technology. Neither Moonshot, Anthropic, nor the White House's own science office would comment when asked.
What makes this notable isn't just the accusation itself, which arrives without any public evidence attached. It's the language. Distillation — training a cheaper model off a pricier frontier model's outputs — is standard practice across the industry, not some exotic crime. By recasting it as theft rather than competition, the administration is opening the door to a much harder response: tighter API controls for U.S. labs, expanded chip export restrictions, maybe both.
The timing adds fuel. Moonshot released Kimi K3 on July 16, a 2.8 trillion-parameter model it billed as the largest open-weight system in the world, with performance it says approaches Fable 5 itself. Full weights weren't even scheduled to drop until July 27. That came roughly five weeks after a U.S. export directive forced Anthropic to briefly pull Fable 5 and Mythos 5 access over national security concerns — access that was only fully restored earlier in July. A rival matching your frontier model's capability that fast, right after your own export lockdown, is exactly the kind of coincidence that gets people in Washington asking questions.
Kratsios also alleges Moonshot got its hands on servers running Nvidia's GB300 chips, deployed in Thailand, likely for training. If true, that points to a harder problem than chip export bans alone can solve: stopping a company from renting compute overseas is a different fight than stopping chips from crossing a border. This isn't happening in isolation either — Anthropic accused Moonshot, DeepSeek and MiniMax back in February of running coordinated distillation campaigns against Claude, tying more than 16 million exchanges across some 24,000 fraudulent accounts to the three firms, with 3.4 million of those linked to Moonshot specifically. The State Department followed with a diplomatic cable in April naming Moonshot directly and warning allies about the practice.
None of this proves anything yet. What it does show is a pattern building toward policy — tighter API rules for U.S. labs, sharper scrutiny of cloud compute rented abroad, maybe sanctions. Whether Kratsios ever produces the technical evidence to back his claim, or whether this becomes just another talking point in a longer trade fight, is the open question everyone in the sector is now sitting with.
My take — AI-written commentary, not fact-checked reporting
An accusation this specific, delivered with zero public evidence, should be treated as exactly that — an accusation, not a finding. Governments have every incentive to dress up ordinary competitive pressure as theft when a rival ships something uncomfortably close to their own frontier model within weeks of an export clampdown. That doesn't mean Moonshot is innocent; it means nobody outside a closed room actually knows yet, and the rhetoric is already outrunning the receipts.}
Read more about this at: The New Stack