Experts Question Claims That Kimi K3 Reached Frontier Level by Copying Anthropic’s Fable

A White House science advisor has accused Chinese AI company Moonshot of building its advanced open-weight model, Kimi K3, by copying Anthropic’s Fable large language model, but multiple AI researchers are pushing back against the narrative that distillation alone explains the model’s capabilities. Michael Kratsios, the White House science advisor, wrote on social media that Moonshot engaged in “large-scale, covert industrial distillation” to steal U.S. technology, while also alleging the company used advanced Nvidia chips banned for export to China. Moonshot did not respond to requests for comment, and Kratsios did not provide evidence for his claims.

Why Experts Are Skeptical of the Distillation Allegation

Distillation is a process in which one AI model is systematically queried to generate data that can train another model. In theory, a company could use this technique to replicate the capabilities of a frontier model like Anthropic’s Fable. However, researchers argue that the timeline and technical requirements make this an unlikely explanation for Kimi K3’s performance. Fable was only publicly released on July 1st. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, told Bitcoin World: “I don’t think you get a model this strong and this quickly on the heels of Fable doing strictly distillation. There’s just not even frankly time, right? You can’t distill that much data, train a model, and release it in two weeks.”

The Technical Limits of Distillation for Frontier Models

Nathan Lambert, an AI researcher at the Allen Institute for AI, said in a recent podcast that distillation has become less impactful over time as Chinese models approach the frontier. He noted that replicating Fable-level capabilities would likely require reinforcement learning techniques, which demand significant infrastructure—potentially tens of millions of agents grading model responses. Running such operations through a frontier lab’s API would be “insanely expensive” and likely a time bottleneck, Lambert said. He added that the benefits of supervised fine-tuning, a simpler form of distillation, are diminishing as models grow more complex.

Broader Context: A History of Distillation Accusations

This is not the first time Anthropic has accused Chinese AI companies of distillation. Earlier this year, Anthropic publicly claimed that Moonshot, DeepSeek, and MiniMax systematically queried its models, using IP addresses and metadata to identify millions of exchanges it said were “distinct from normal usage patterns.” Anthropic did not respond to Bitcoin World’s queries about Fable specifically. However, the practice is not unique to China. Elon Musk testified earlier this year that his company SpaceXAI distilled OpenAI models to develop Grok, calling the practice common in the industry. The line between distillation and legitimate synthetic data generation can be blurry.

Chinese Technical Expertise vs. the Narrative of Copying

Experts caution against underestimating the capabilities of Chinese AI teams. Hancock noted that one of Moonshot’s founders was a PhD student at Carnegie Mellon University. “These are legitimate researchers and engineers doing solid work,” he said. “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.” This perspective challenges the framing that Chinese models are solely derivative, pointing instead to independent innovation.

The Chip Export Issue: A Separate but Connected Concern

Kratsios’s allegations also included claims that Moonshot obtained advanced Nvidia Grace Blackwell 300 chips, which are banned for export to China, and accessed GB300-equipped servers in Thailand. A black market for such chips exists, according to Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology. In May, the founder of Supermicro, a U.S. server builder, was indicted for smuggling advanced chips into China. Bresnick advocated for “know your customer” laws for data centers globally, noting that proposed federal rules from the Biden administration in 2024 have not advanced under President Trump.

Conclusion

The debate over Kimi K3’s origins highlights a broader tension in the AI industry: how to distinguish between legitimate competitive progress and intellectual property theft. While the White House has framed the issue as a clear case of industrial espionage, technical experts point to significant gaps in that narrative. The lack of public evidence from Kratsios, combined with the technical and timeline constraints, suggests that the story may be more complex than official accusations imply. As the U.S. and China continue to compete in AI, the line between copying and independent innovation will remain a flashpoint.

FAQs

Q1: What is AI distillation?
Distillation is a process where one AI model is systematically queried to generate training data for another model, potentially allowing the second model to replicate some of the first model’s capabilities.

Q2: Why do experts doubt that Kimi K3 was built solely by distilling Anthropic’s Fable?
Because Fable was only released in July, leaving insufficient time for large-scale distillation, training, and release. Additionally, frontier-level capabilities likely require advanced reinforcement learning techniques that are costly and time-prohibitive via API.

Q3: What are the chip export allegations against Moonshot?
White House advisor Michael Kratsios claimed Moonshot obtained advanced Nvidia Grace Blackwell 300 chips, which are banned for export to China, and accessed servers in Thailand equipped with those chips. A black market for such chips is known to exist.