The White House science advisor made a serious accusation last week. Michael Kratsios claimed that Moonshot AI, the Chinese company behind Kimi K3, built its powerful open-weight model by copying Anthropic’s Fable LLM and used chips that are banned from export to China. ‘Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable,’ Kratsios wrote on social media.
His comments came as U.S. officials are reportedly discussing whether to ban Chinese open-weight models entirely, a move that has unsettled parts of the AI industry. Treasury Secretary Scott Bessent added fuel to the fire, saying the U.S. is finding ‘watermarks of our U.S. large language models on many of the Chinese models.’ Neither the Treasury Department nor Moonshot responded to requests for comment or detail.
But according to TechCrunch, AI researchers who study these systems closely are not buying the core of the argument. The timeline alone, they say, makes the distillation claim hard to believe.
Distillation is the process of systematically querying an existing AI model to understand how it works and copy its capabilities into a new one. It can involve asking the model to show its reasoning step by step, or using its outputs to train a separate model through a process called supervised fine-tuning, or SFT. This is also why models trained this way sometimes identify themselves as Claude, the name of Anthropic’s assistant, even when they’re not. As researcher Nathan Lambert put it, fine-tuning is ‘where the model picks up its manners.’
The problem with the Fable distillation theory is simple math. Fable only became publicly available on July 1st. Kimi K3 followed just two weeks later. ‘You can’t distill that much data, train a model, and release it in two weeks,’ said Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI. ‘There’s just not even frankly time.’
Hancock is not alone in that view. Lambert, an AI researcher at the Allen Institute for AI, argued in a recent podcast that distillation is becoming less effective anyway as Chinese models close the gap with Western ones. His point: if distilling Kimi K3’s outputs were enough to replicate its performance, other labs would already be doing it. They’re not seeing those results.
There’s also a practical cost problem with trying to distill a frontier model at the scale needed to match something like Kimi K3. The most advanced techniques require reinforcement learning, which means running tens of millions of agents to grade and improve a model’s responses. Doing that through a commercial API would be, in Lambert’s words, ‘insanely expensive’ and almost certainly too slow to produce results at this speed.
That doesn’t mean distillation played no role. Anthropic accused Moonshot, along with DeepSeek and MiniMax, of systematically querying its models earlier this year. The company said it identified millions of exchanges linked to those companies through IP addresses and other metadata, describing the queries as ‘distinct from normal usage patterns, reflecting deliberate capability extraction rather than legitimate use.’ Anthropic did not respond to questions about whether Fable specifically was distilled.
It’s also worth noting that distillation is not unique to Chinese AI labs. Elon Musk testified earlier this year that SpaceXAI distilled OpenAI models to help build Grok, and described the practice as common across the industry. The line between distillation and building synthetic training datasets is often blurry, and many Western labs have used outputs from other models in their training pipelines.
‘In general, Americans are understating the technical expertise of these Chinese teams,’ Hancock said. ‘One of the founders of Moonshot was a CMU PhD student. These are legitimate researchers and engineers doing solid work. If American models ground to a halt, I think China’s progress would slow, but would still continue. They’re not just riding coattails here.’
The chip access question is harder to dismiss. Kratsios also alleged that Moonshot obtained Nvidia Grace Blackwell GB300 chips, which are banned from export to China, and used GB300-equipped servers in Thailand. A black market for restricted chips does exist, according to Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology. In May, the founder of Supermicro was indicted for smuggling advanced chips into China.
Bresnick wants stronger accountability at the data center level:
- Know-your-customer rules for data centers worldwide
- Reporting requirements for companies conducting large training runs on high-end hardware
- Clearer enforcement of existing export controls that require sellers to verify end use
The Biden administration proposed federal know-your-customer rules for data centers in 2024, but no follow-through has happened under Trump. For now, exporters are technically required to ensure their chips are only used for approved purposes, but enforcement remains patchy at best.
