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

What the report says

TechCrunch reports that U.S. officials have accused Moonshot, the Chinese company behind the open-weight Kimi K3 language model, of relying on Anthropic’s Fable model and restricted Nvidia hardware. White House science adviser Michael Kratsios alleged that Moonshot copied U.S. technology through large-scale distillation, while Treasury Secretary Scott Bessent separately said U.S. “watermarks” were being found in Chinese models. TechCrunch noted that neither official provided detailed evidence in response to questions, and Moonshot did not answer inquiries about how Kimi K3 was trained.

Several AI researchers cited by TechCrunch were skeptical that distillation from Fable alone could explain Kimi K3’s performance, especially given the short time since Fable became publicly available. Braden Hancock of the Laude Institute and Snorkel AI said the timeline made a full distillation-based explanation unlikely. Nathan Lambert of the Allen Institute for AI argued that simple supervised fine-tuning is becoming less decisive as frontier models increasingly depend on reinforcement learning and extensive infrastructure.

TechCrunch also reported that Anthropic had previously accused Moonshot, DeepSeek and MiniMax of systematically extracting capabilities from its models, based on usage patterns it described as abnormal. But the article emphasized that distillation and synthetic-data practices are widespread in the AI industry and can be difficult to distinguish.

The report also examined the hardware side of Kratsios’ claim, including alleged access to Nvidia Grace Blackwell chips or servers outside China. Sam Bresnick of Georgetown’s Center for Security and Emerging Technology said chip smuggling markets exist and supported stronger “know your customer” rules for global data centers. The debate matters because it connects AI competition, export controls and possible restrictions on Chinese open-weight models.