White House science adviser Michael Kratsios has accused Moonshot AI, the Chinese company behind the open-weight Kimi K3 language model, of copying Anthropic’s Fable model. He also alleged that the company used advanced computing hardware that was not cleared for export to China.
Kratsios characterized the alleged activity as large-scale industrial distillation intended to extract proprietary American technology. He did not disclose the evidence behind the accusations or provide details about how the alleged copying was identified. Moonshot did not respond to questions about its training process.
The claims arrive amid a broader policy dispute over Chinese open-weight models and whether the United States should impose additional restrictions on their distribution or use. But researchers who study model training are skeptical that distillation from Fable could, by itself, account for Kimi K3’s performance.
Why researchers question the distillation timeline
Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, argued that Kimi K3 appeared too quickly for a straightforward Fable distillation operation to explain the finished model. His assessment does not rule out the possibility that Moonshot used outputs from other models. It challenges the narrower idea that Fable was the primary reason Kimi K3 became so capable.
Building a model through distillation is more involved than copying a set of answers. In the form discussed by AI researchers, the process generally involves querying a more capable model, collecting useful outputs and using that material during another model’s post-training. One possible approach uses prompts and responses as supervised fine-tuning data, though the results depend heavily on the quality and scale of the training process.
Nathan Lambert, an AI researcher at the Allen Institute for AI, has argued that supervised fine-tuning alone becomes less decisive as models move closer to the technical frontier. In his view, fine-tuning can shape how a model responds and behaves without necessarily transferring the deeper reasoning capabilities responsible for strong performance.
Lambert suggested that reproducing more advanced capabilities might require reinforcement-learning methods rather than supervised fine-tuning by itself. That remains an expert assessment rather than a confirmed account of how Kimi K3 was trained. It also points to a practical constraint: large reinforcement-learning runs can require substantial computing capacity, many automated evaluation agents and repeated rounds of grading and adjustment.
Attempting that process through a frontier model provider’s public API could be expensive and slow. It might also fail to deliver a meaningful performance gain. Those limitations make a rapid, Fable-centered explanation less convincing to the researchers questioning Kratsios’ allegation.
Earlier models could still have played a role
Skepticism about the Fable claim does not establish that Moonshot trained Kimi K3 without assistance from other models. Anthropic previously accused Moonshot, DeepSeek and MiniMax of systematically extracting capabilities from its systems. The company said it identified millions of exchanges associated with users it linked to those organizations through IP addresses and other metadata.
Anthropic characterized those interactions as capability-extraction activity rather than ordinary use. The allegation does not, however, reveal the precise role any collected outputs may have played in a finished model. It also does not establish that Fable supplied the capabilities responsible for Kimi K3’s performance.
The distinction is important because model development can combine pretraining, synthetic data, fine-tuning, reinforcement learning and outputs from multiple systems. Seeing traces of another model’s response style would not automatically show that the other model supplied the underlying architecture, training data or full set of capabilities.
Hancock also cautioned against treating Chinese AI teams as technically dependent on American labs. He described Moonshot’s researchers and engineers as capable of making substantial progress through their own work. His assessment was that Chinese AI development could slow if access to leading American models disappeared, but it would not simply stop.
That argument complicates the political framing around Kimi K3. Distillation may be one ingredient in a training pipeline without being the central explanation for the model’s strength. Treating every competitive Chinese model as a direct copy risks overlooking the research, infrastructure and engineering needed to build and operate it.
The separate dispute over advanced AI chips
Kratsios paired the distillation accusation with another allegation: that Moonshot obtained Nvidia Grace Blackwell 300 chips and accessed servers equipped with GB300 hardware in Thailand. The claim has not been independently substantiated, and the details of the alleged access remain unclear.
Advanced AI chip export controls are designed to limit China’s access to hardware that can support large training runs. Sam Bresnick, a research fellow at Georgetown University’s Center for Security and Emerging Technology, has said an illicit market for restricted chips exists. He has advocated for know-your-customer requirements that would oblige data centers to identify organizations conducting unusually large training jobs on state-of-the-art hardware.
The concern is that remote access can weaken controls focused mainly on the physical destination of a chip. A company may not need to import restricted hardware directly if it can rent sufficient computing capacity from servers in another country. That possibility shifts part of the enforcement problem toward cloud providers and data-center operators.
Scrutiny of hardware supply chains has also grown following an indictment involving a founder of US server manufacturer Supermicro, who was accused in May of smuggling advanced chips into China. An indictment is an allegation, not a finding of guilt, but the case illustrates why policymakers are looking beyond manufacturers when considering enforcement.
The Commerce Department under President Joe Biden proposed federal know-your-customer rules for data centers in 2024. No comparable expansion described in the proposal appears to have advanced under President Donald Trump. Bresnick argues that operators enabling major training runs should face reporting requirements covering who is using the hardware and for what purpose.
Kimi K3 leaves two questions, not one
The controversy ultimately combines two issues that should be evaluated separately. The first is whether Moonshot used outputs from Anthropic or other frontier models during Kimi K3’s development. The second is whether the company obtained or remotely accessed advanced hardware covered by US export restrictions.
Neither question is answered merely by Kimi K3’s strong performance. Researchers challenging the Fable theory are not claiming that distillation never occurred. Their narrower point is that the timing, scale and technical demands make it unlikely that Fable distillation alone produced the model’s capabilities.
Without additional evidence about Moonshot’s training data, reinforcement-learning pipeline and computing infrastructure, the strongest accusations remain allegations. What the debate already shows is that model provenance is becoming harder to establish as AI developers combine public research, synthetic data, model-generated outputs and increasingly complex post-training systems.