Experts Question White House Claim That Kimi K3 Was Built by Copying Anthropic Models

White House science adviser Michael Kratsios says Moonshot’s Kimi K3 was built by copying Anthropic’s Fable model and training on restricted Nvidia chips, but AI researchers say that explanation does not fully account for how the Chinese model became so capable so quickly. The dispute matters because it sits at the intersection of AI model theft allegations, chip export controls, and the growing competition between U.S. and Chinese frontier labs.

The controversy erupted after Kratsios accused Moonshot, the Beijing-based company behind Kimi K3, of engaging in “large-scale, covert industrial distillation” aimed at stealing American technology. His comments came as the U.S. government has been weighing tougher restrictions on Chinese open-weight models, while the AI industry continues to debate how much model performance can really be reproduced through copying alone.

But researchers who study model distillation say the timeline and the technical demands make the accusation incomplete at best. Their view: distillation may have played some role, as it often does across the AI industry, but Kimi K3’s strength likely reflects a broader and more expensive training pipeline that goes beyond simply querying Anthropic’s system.

What sparked the latest dispute over Kimi K3?

The dispute began with public comments from Kratsios, who suggested that Moonshot had copied Anthropic’s Fable model and combined that effort with advanced Nvidia hardware that is not supposed to be exported to China. His remarks echoed earlier criticism from Treasury Secretary Scott Bessent, who said the U.S. has found traces of American models inside Chinese systems and called that outcome unacceptable.

Kratsios did not provide evidence supporting the allegation, and Moonshot did not respond to questions about how Kimi K3 was trained. The White House adviser also did not explain what information led him to his conclusion, leaving observers to infer that the claim was based on a mix of technical forensics, intelligence, and broader concern about Chinese AI progress.

The timing is notable. Anthropic’s Fable model was publicly released only on July 1, and Kimi K3 has emerged soon after as one of the most prominent open-weight large language models available. That rapid rise has fueled speculation that Moonshot may have benefited from access to frontier-model outputs in ways that could accelerate performance.

How does AI model distillation actually work?

Distillation is a training method in which one model is used to help build another. In practice, a company repeatedly queries a target model, collects the answers, and uses those outputs to train a new system that imitates some of the target’s capabilities, style, or decision patterns.

The simplest version resembles ordinary supervised fine-tuning, where prompt-response pairs teach a smaller model to behave more like a larger one. In more advanced forms, a lab may also ask the target system to reveal reasoning steps or chain-of-thought-like explanations, then reuse those traces to improve a new model’s performance.

That is why distillation has long been seen as a shortcut in AI development. It can help a new model learn tone, formatting, and some task behavior quickly, especially when a company wants a product that feels similar to an established chatbot.

“Fine tuning is where the model picks up its manners,” Nathan Lambert, an AI researcher at the Allen Institute for AI, said, summarizing how imitation can shape a system’s personality without necessarily recreating the full depth of the original.

Even so, researchers stress that style transfer is not the same thing as reproducing frontier capability. A system may imitate a model’s voice or answer structure without matching its reasoning depth, robustness, or tool-use quality.

Why experts say distillation alone is unlikely to explain Kimi K3

Experts interviewed about the allegations say Kimi K3 appears too capable, and appears to have arrived too quickly, to be explained by distillation from Fable alone. Their argument is built on both timing and compute.

Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, said the timeline does not make sense for a model of this scale. Fable was publicly available only in early July, leaving only a narrow window for Moonshot to query it, build a training set, run post-training, and release a highly competitive system.

“I don’t think you get a model this strong and this quickly on the heels of Fable doing strictly distillation,” Hancock said. “There’s just not even frankly time.”

Lambert has made a similar point in public commentary, arguing that the role of simple supervised fine-tuning is shrinking as models become more sophisticated and as competitive labs increasingly rely on reinforcement learning.

“Distillation has become less and less impactful over time as the Chinese models get closer to the frontier and the training regime shifts to reinforcement learning,” Lambert said in a recent podcast.

That distinction matters. If a model is merely mimicking responses, it may absorb surface-level patterns. But if a system is genuinely near the frontier, its creators likely needed stronger training signals, broader data pipelines, and substantial reinforcement learning infrastructure.

Why reinforcement learning changes the picture

Reinforcement learning changes the picture because it can be used to shape behavior at a much deeper level than basic imitation. Instead of copying one model’s answers, developers can score a smaller model’s outputs and refine it through repeated trial and error.

According to Lambert, if a lab wanted to reproduce Fable-like strengths, it would likely need these more advanced methods rather than plain supervised fine-tuning. In those systems, the larger model can act as a judge, rating the smaller model’s responses and creating a training loop that improves reasoning over time.

That process is not just technically harder. It is also far more expensive and demanding in terms of infrastructure, compute orchestration, and throughput.

Lambert argued that using an API from a frontier lab to perform that scale of training would be costly and potentially too slow to make a viable shortcut. Models that are already slow to query can become a bottleneck when millions of evaluations are needed.

What did Anthropic previously allege about Moonshot and other Chinese labs?

Anthropic has already publicly accused Moonshot, DeepSeek, and MiniMax of systematic distillation earlier this year. The company said it detected millions of exchanges between its own systems and users linked to those firms through IP addresses and other metadata.

Anthropic said the query patterns were not consistent with ordinary consumer use. Instead, it described them as deliberate attempts to extract capabilities from its models for competitive development.

That accusation is important because it suggests the broader issue is not limited to one Chinese company or one model. Rather, it reflects a recurring concern among frontier labs that their outputs can be harvested at scale and reused as training fuel by competitors.

Anthropic did not respond to questions about whether the Fable model itself was targeted in the same way, and Moonshot has not publicly addressed the specific allegation that Kimi K3 was built through Fable distillation.

How common is distillation across the AI industry?

Distillation is common across the AI industry, and not only in China. The practice sits in a gray area between legitimate benchmarking, competitive product development, and outright appropriation of another company’s model behavior.

Elon Musk testified earlier this year that his company SpaceXAI had used OpenAI models in the process of developing Grok, saying the approach was widespread in the sector. That testimony underscored how difficult it is to draw a clean line between inspiration, imitation, and reverse engineering.

The boundaries are especially blurry because some companies also generate synthetic datasets that resemble distillation outputs without necessarily depending on direct copying. In practice, many teams use a combination of model-generated examples, human annotation, and self-play-style training to improve performance.

That is why researchers caution against treating every high-performing model as proof of illicit copying. Distillation can explain certain behaviors, but it rarely tells the whole story of a frontier model’s development.

Did Moonshot also get access to banned Nvidia chips?

That appears to be the other half of the allegation, and it may matter as much as the distillation claim. Kratsios also suggested Moonshot used advanced Nvidia hardware that should not have been available for export to China, including Grace Blackwell 300 chips and GB300-equipped servers in Thailand.

Those components are among the most advanced AI accelerators in the world, and they are restricted under U.S. export rules. If Chinese firms gained access through overseas intermediaries, the issue would point not only to model-copying concerns but also to enforcement gaps in the global chip supply chain.

Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology, said a black market for advanced chips does exist. He pointed to the broader need for stronger controls on who is allowed to train on top-tier hardware and for what purpose.

“If you are letting a company conduct huge training runs on your state-of-the-art hardware, there needs to be a reporting mechanism for who that company is and what they’re doing,” Bresnick said.

The U.S. has already seen criminal cases tied to chip smuggling. In May, the founder of Supermicro was indicted over allegations involving the movement of advanced chips into China, underscoring how enforcement problems can arise long after export rules are written.

What are know-your-customer rules for data centers?

Know-your-customer rules would require data centers and infrastructure providers to identify who is renting advanced AI capacity and what kind of work they are running. The idea is to make it harder for prohibited entities to hide behind shell companies or third-party intermediaries.

Such rules were proposed by President Joe Biden’s Commerce Department in 2024, but there has been no major follow-through under the Trump administration, according to the reporting. For policymakers worried about both model theft and hardware diversion, that stalled effort remains a missed opportunity.

Without tighter oversight, exporters are left to ensure advanced chips are used only for permitted purposes, a standard that is difficult to verify once hardware crosses borders or changes hands through brokers.

What do researchers think Moonshot’s progress really means?

Researchers say Moonshot’s success should not be reduced to a story of theft alone. Hancock argued that U.S. observers often underestimate the technical skill inside Chinese AI firms, including Moonshot.

“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.”

His broader point is that Chinese labs are not merely borrowing from U.S. models and sitting still. Even if American labs suddenly slowed, Chinese teams would likely continue progressing because they have their own talent pipelines, research culture, and access to a large domestic market.

That view complicates the political narrative around model theft. Yes, copying may happen. Yes, export controls matter. But it does not follow that every competitive Chinese model is simply an American model in disguise.

Why the debate matters beyond one model

The Kimi K3 controversy reaches far beyond Moonshot and Anthropic. It touches the strategic rivalry over AI leadership, the reliability of open-weight model releases, the limits of export controls, and the practical difficulty of proving that a model was built through illegitimate means.

For U.S. policymakers, the issue is twofold: stopping sensitive technology from moving through restricted channels and protecting the competitive advantage of domestic labs whose outputs can be mined for training data. For AI companies, the concern is that their models may be used as silent teachers for rivals who never pay for a license or disclose the source of their gains.

For researchers, the debate is a reminder that “distillation” is not a single technique but a spectrum of practices. Some are routine and lawful. Others can be evasive, industrial-scale, and designed specifically to extract proprietary capabilities.

The challenge for regulators is that the technical evidence is often hard to interpret. A model may show signs of imitation, but imitation alone does not prove theft. Hardware access may be suspicious, but suspicious access does not by itself reveal how a system was trained.

Key facts at a glance

Timeline of the dispute

What happens next?

For now, the evidence in public remains incomplete. Kratsios has made a serious accusation, Treasury has echoed the broader concern, and Anthropic has already claimed that Chinese rivals have repeatedly probed its systems. But the experts most familiar with model training say the available facts do not support a simple story in which Kimi K3 was produced by a quick imitation of Fable.

The more likely explanation, according to those experts, is a mixed one: some level of distillation or model reuse, a larger and more sophisticated training program, and possible access to restricted infrastructure that could have accelerated development. If that is right, then the controversy is not only about one alleged theft. It is about the difficulty of policing an AI race in which software, data, and hardware all cross borders in opaque ways.

That makes Kimi K3 less of a single-model scandal than a case study in the new geopolitics of AI. The fight is no longer only over who builds the best systems. It is also over who can use whose models, whose chips can be moved where, and how any of it can be proven after the fact.

“If American models ground to a halt, I think China’s progress would slow, but would still continue,” Hancock said. “They’re not just riding coattails here.”

That is the central takeaway from the Kimi K3 debate: copying may be part of the story, but it is not the whole story, and the real competitive gap is being shaped by talent, infrastructure, and policy as much as by model outputs.