White House Claims Moonshot Copied Anthropic's Fable to Build Kimi K3, but Researchers Say the Math Doesn't Work

Kimi K3 Copied From Anthropic's Fable? Researchers Say the White House's Math Doesn't Work

Moonshot, the Chinese company that developed Kimi K3, the largest available open-weight LLM, is facing an accusation from the highest level of the U.S. government: that it built the model by distilling Anthropic's Fable on export-controlled Nvidia chips. Researchers who study how models actually learn say the timeline and the economics make that claim nearly impossible to support.

The problem comes down to two dates. Fable has been publicly available since July 1. Kimi K3 arrived roughly two weeks later 1. In that window, no lab can generate enough training data through a competitor's API, run reinforcement learning at the required scale, and ship a frontier-grade model. That is the case three researchers made to TechCrunch, and it rests on three technical constraints.

White House science advisor Michael Kratsios condemned what he called "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology" as the administration reportedly weighs banning Chinese open-weight models 1. Treasury Secretary Scott Bessent separately claimed officials are finding "watermarks" of U.S. large language models on Chinese models, though it remains unclear what those watermarks consist of 1. Moonshot has not responded to questions about how it trains its models, and Kratsios offered no evidence or further detail to support his claims 1.

The first constraint is time. "You can't distill that much data, train a model, and release it in two weeks," Braden Hancock, who co-founded Snorkel AI and is now a researcher at the Laude Institute, told TechCrunch 1. Distillation works by systematically querying a target model to map its behavior, then using the responses to train a new model. Hancock's point is straightforward: Fable has not existed long enough to supply the volume of data Kimi K3 would require.

The second is diminishing returns. Nathan Lambert, an AI researcher at the Allen Institute for AI, said distillation has become less impactful as Chinese models close the gap with the frontier and training shifts toward reinforcement learning 1.

The third is cost. Replicating Fable-level capabilities through distillation would likely require reinforcement learning, where agents of the larger model grade and adjust a smaller model's responses. Those RL runs can involve tens of millions of agents. Routing that volume of work through a competitor's paid API would be, in Lambert's assessment, "insanely expensive," and the models' response times would create a bottleneck that might not even yield a performance gain 1.

None of this means distillation is not happening. Anthropic publicly accused Moonshot, DeepSeek, and MiniMax of systematically distilling its models earlier this year, saying it detected millions of queries from those companies' networks that did not match normal usage patterns and appeared designed to extract capabilities 1. Elon Musk testified that SpaceXAI distilled OpenAI models to develop Grok and called the practice common across the industry 1. The boundary between distillation and synthetic data development is fuzzy, and the practice is not unique to China.

What researchers are saying is narrower and more specific: distillation may have contributed to Kimi K3's development, but it does not explain the model's capabilities on the timeline alleged. "Americans are understating the technical expertise of these Chinese teams," Hancock said. "These are legitimate researchers and engineers doing solid work" 1. One Moonshot co-founder trained at Carnegie Mellon. "If American models ground to a halt, I think China's progress would slow, but would still continue," Hancock added. "They're not just riding coattails here" 1.

The second prong of Kratsios's claim concerns hardware. He said Moonshot obtained Nvidia Grace Blackwell 300 chips, also known as GB300s, and accessed GB300-equipped servers in Thailand 1. Those chips are banned from export to China. Sam Bresnick, a research fellow at Georgetown's Center for Security and Emerging Technology, said a black market for advanced chips exists 1.

Bresnick backs "know your customer" requirements at data centers worldwide, forcing operators to disclose who is running large training workloads and on what hardware 1. The Biden administration's Commerce Department proposed federal rules along those lines in 2024. No further progress appears to have followed under the current administration 1.

The gap between Kratsios's accusation and the researchers' pushback is where the real story sits. The U.S. is preparing its next round of AI export controls on a theory of how models are built that does not survive contact with the people who build them. When a presidential advisor says systematic distillation and researchers say two weeks is not enough time, the policy constructed on the first claim will miss its target. The coming restrictions on Chinese AI will be shaped by a political framing of a technical process, and the experts who understand that process are already saying the framing is wrong.