The White House Says China Cloned Claude to Build Kimi K3. There Wasn't Enough Time, Researchers Say.

Six days after Kimi K3 became the largest open-weight AI model ever released, the White House accused its maker of stealing the model that made it possible.

On July 22, Michael Kratsios, director of the White House Office of Science and Technology Policy, posted on X that Beijing-based Moonshot AI had "distilled Anthropic's Fable for the development of its K3 model," using what he described as "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." He added that Moonshot had "acquired GB300-equipped servers and has accessed GB300s in Thailand, likely to train its AI models" — Nvidia's most advanced Blackwell-generation chip, and one the U.S. has restricted from export to China for years. Kratsios called the alleged campaign "unacceptable."

It's a serious charge, and it landed at a delicate moment. Moonshot is reportedly preparing a final pre-IPO funding round in August, targeting a $50 billion valuation ahead of a Hong Kong listing that could arrive before year's end, according to Bloomberg. Kimi K3 — a 2.8 trillion-parameter mixture-of-experts model with a 1-million-token context window — is the reason investors are interested in the first place. Moonshot says it's the largest open-weight model ever built, and it has already posted results beating Claude Fable 5 on at least one closely watched coding benchmark, Frontend Code Arena, per Tom's Hardware.

THE TIMELINE PROBLEM

Here's what makes independent AI researchers skeptical: Fable 5 only became available to the general public on July 1. It had actually launched June 9, but the U.S. government applied export controls to it three days later over national-security concerns about foreign access, forcing Anthropic to suspend the model entirely until the restrictions lifted on June 30. Kimi K3 shipped on July 16 — fifteen days after Fable 5 was even accessible to distill from.

"I don't think you get a model this strong and this quickly on the heels of Fable doing strictly distillation," Braden Hancock, co-founder of Snorkel AI and a researcher at the Laude Institute, told TechCrunch. Distillation — training a smaller or competing model by learning from a larger model's outputs — is a real and widely used technique, but doing it at a scale sufficient to produce a 2.8-trillion-parameter frontier model, then training and shipping that model, in barely two weeks strikes Hancock as implausible. He also pushed back on the framing itself, noting Moonshot's team includes CMU-trained researchers: "These are legitimate researchers and engineers doing solid work."

Nathan Lambert of the Allen Institute for AI raised a related, more technical objection: the reinforcement-learning-based distillation techniques capable of meaningfully lifting a frontier model's performance require enormous compute — potentially tens of millions of agent rollouts — and running that against a rival lab's paid API would be, in his words, "insanely expensive," with "time bottlenecks" that would blow past any two-week window. Simpler supervised fine-tuning on a competitor's outputs, meanwhile, delivers diminishing returns the closer a model gets to the frontier. If SFT-on-outputs were really enough to conjure a K3-caliber model out of Fable, Lambert argued, "everyone would be easily able to catch up" the same way.

Kratsios did not publish evidence for the distillation claim, and Moonshot did not respond to requests for comment from multiple outlets that covered the accusation.

THE CHIP TRAIL IS A SEPARATE — AND HARDER TO WAVE AWAY — STORY

The Nvidia allegation stands on different footing. Thailand isn't covered by the same export restrictions that bar advanced chips from reaching China directly, and data-center capacity in Southeast Asia has grown quickly as hyperscalers diversify where they build. That gap is exactly the kind of enforcement seam sanctions-watchers have been warning about for months: a lab doesn't need to smuggle a physical chip across a border if it can rent compute sitting on the other side of one. Whether Moonshot actually did what Kratsios describes hasn't been independently verified either, but it's a narrower, more checkable claim than "they cloned our model" — and it's the one likely to draw the most scrutiny from export-control enforcers rather than AI researchers.

A PATTERN, NOT AN ISOLATED INCIDENT

This isn't the first time Washington and Anthropic have leveled this kind of accusation at a Chinese lab. In a June 10 letter to the Senate Banking Committee, Anthropic accused Alibaba's Qwen team of running what it called the largest known distillation campaign against Claude: roughly 25,000 fraudulent accounts generating more than 28.8 million exchanges with Claude between April 22 and June 5, allegedly to harvest outputs for training Qwen's models. Alibaba denies it. Neither figure in either case — Alibaba's 28.8 million queries or Moonshot's alleged distillation platform — has been independently verified by an outside auditor; both rest on the accusing party's own telemetry and analysis.

Congress has been circling this territory since at least April, when House committees opened investigations into how thoroughly Chinese AI labs are integrating U.S. model outputs into their own training pipelines. The Moonshot accusation slots into that same investigation, and into a broader pattern: as Chinese open-weight models close the performance gap with U.S. frontier labs faster than many expected, the U.S. government's go-to explanation is increasingly "they copied it," while the researchers who actually build these systems for a living are increasingly unconvinced the math supports that story on this particular timeline.

WHAT TO WATCH

Two threads here are worth separating, because they'll likely resolve on different clocks: whether investigators ever produce evidence specific enough to convince skeptical researchers that distillation explains Kimi K3, and whether the Thailand chip-access allegation turns into an actual enforcement action against the compute providers involved. The second one doesn't require anyone to agree on what happened inside Moonshot's training runs — it just requires proving where the GPUs physically were. That's a much easier case to make, and a much easier one for Washington to actually win.

Sources

Michael Kratsios statement on X: https://x.com/mkratsios47/status/2079933645888880708

CyberScoop, White House accuses Moonshot AI of Anthropic model distillation: https://cyberscoop.com/white-house-accuses-moonshot-ai-anthropic-model-distillation/

TechCrunch, Experts say exploiting Anthropic's Fable isn't how Kimi K3 got so good: https://techcrunch.com/2026/07/23/experts-say-exploiting-anthropics-fable-isnt-how-kimi-k3-got-so-good/

Bloomberg, China's Moonshot in Talks on Pre-IPO Funds at $50 Billion Value: https://www.bloomberg.com/news/articles/2026-07-21/china-s-moonshot-in-talks-on-pre-ipo-funds-at-50-billion-value

TechTimes, Alibaba Ran Largest Known AI Theft Campaign Against Claude, Anthropic Tells Senate: https://www.techtimes.com/articles/319105/20260625/alibaba-ran-largest-known-ai-theft-campaign-against-claude-anthropic-tells-senate.htm

Gizmochina, Kimi K3: Moonshot AI unleashes 2.8 trillion parameter model for free: https://www.gizmochina.com/2026/07/19/kimi-k3-moonshot-ai-unleashes-2-8-trillion-parameter-model-for-free/

Tom's Hardware, China's 2.8-trillion-parameter Kimi K3 beats Claude Fable 5 in Frontend Code Arena benchmark: https://www.tomshardware.com/tech-industry/artificial-intelligence/moonshot-releases-2-8-trillion-parameter-kimi-k3

Anthropic Newsroom, Redeploying Claude Fable 5: https://www.anthropic.com/news/redeploying-fable-5