Experts say exploiting Anthropic’s Fable isn’t how Kimi K3 got so good

US Accusations of IP Theft and Export Violations

White House science advisor Michael Kratsios and Treasury Secretary Scott Bessent have leveled serious accusations against Moonshot, the Chinese company behind the Kimi K3, one of the largest available open-weight large language models. Kratsios claimed that Moonshot built Kimi K3 by copying Anthropic’s Fable LLM and utilized advanced Nvidia chips, specifically Grace Blackwell 300s, which are banned from export to China. He characterized this as “large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research.”

These allegations are part of a broader concern within the U.S. government, with Bessent noting the presence of “watermarks of our U.S. large language models on many of the Chinese models.” The claims have fueled discussions about potentially banning Chinese open-weight models, signaling an escalation in the ongoing tech rivalry between the two nations. Moonshot has not publicly responded to these specific allegations, and U.S. officials have not provided detailed evidence to support their claims of direct IP theft.

Expert Skepticism on Distillation's Role

Despite the strong accusations from U.S. officials, AI experts express significant skepticism that simple distillation is the primary method behind Kimi K3's advanced capabilities. Braden Hancock, a researcher at the Laude Institute and co-founder of Snorkel AI, pointed out the impracticality, stating, “I don’t think you get a model this strong and this quickly on the heels of Fable doing strictly distillation.” He highlighted the tight timeline, noting Fable’s public availability only two weeks prior to Kimi K3’s release, making extensive data distillation and training implausible within that timeframe.

Nathan Lambert, an AI researcher at the Allen Institute for AI, echoed this sentiment, suggesting that distillation has become less impactful as Chinese models approach the frontier of AI development and training regimes shift towards reinforcement learning. He explained that while supervised fine-tuning (SFT) can involve using prompts and responses from a target model to train a new one, the benefits of SFT alone are diminishing for complex, frontier models. Achieving Fable-like capabilities would likely necessitate sophisticated reinforcement learning techniques, which involve agents grading responses and require immense computational infrastructure, making API-based distillation prohibitively expensive and slow.

While Anthropic previously accused Moonshot, DeepSeek, and MiniMax of systematically distilling its models earlier in the year, detecting “distinct from normal usage patterns” through IP addresses and metadata, the current debate centers on whether such methods could yield a model as advanced as Kimi K3 so rapidly. Experts suggest that previous frontier models might have contributed, but the rapid leap implies more than just basic distillation.

Broader Implications for AI Development and Regulation

The debate surrounding Kimi K3 underscores the blurry lines in AI development, particularly between distillation and the creation of synthetic datasets. Elon Musk’s testimony that SpaceXAI distilled OpenAI models for Grok highlights that this practice is common across the industry, not just in China. Experts like Hancock emphasize that American officials may be underestimating the technical expertise of Chinese teams, many of whom are legitimate researchers and engineers capable of solid independent work. This suggests that even if American models were to halt progress, China’s AI development would likely continue, albeit at a slower pace.

The second part of the U.S. allegations, concerning Moonshot’s access to banned Nvidia Grace Blackwell 300 chips and GB300-equipped servers in Thailand, points to significant challenges in enforcing export controls. Sam Bresnick, a research fellow at Georgetown’s Center for Security and Emerging Technology, confirmed the existence of a black market for these chips and advocated for “know your customer” laws for data centers globally. He argued for reporting mechanisms to identify companies conducting large training runs on state-of-the-art hardware. While the Biden administration proposed such rules in 2024, progress appears to have stalled under the subsequent administration, leaving a critical loophole in the enforcement of chip export restrictions and raising concerns about the integrity of global AI supply chains and national security.