Kimi K3's capabilities likely not from distilling Anthropic's Fable, experts say
Original: Experts say exploiting Anthropic’s Fable isn’t how Kimi K3 got so good
Why This Matters
The debate highlights growing U.S.-China AI tensions and the technical limits of distillation as a competitive shortcut.
White House science advisor Michael Kratsios claimed Chinese AI firm Moonshot built Kimi K3 by distilling Anthropic's Fable LLM using restricted chips. But AI researchers argue the timeline and technical evidence make large-scale distillation an unlikely explanation for the model's advanced performance.
White House science advisor Michael Kratsios publicly accused Moonshot, the Chinese company behind the open-weight model Kimi K3, of building the model through 'large-scale, covert industrial distillation' of Anthropic's Fable LLM, while also using export-restricted chips. Treasury Secretary Scott Bessent similarly claimed that U.S. LLM 'watermarks' have been found on Chinese models. However, neither official provided specific evidence, and Moonshot did not respond to requests for comment.
AI experts are skeptical. Braden Hancock of the Laude Institute and Snorkel AI told TechCrunch that the timeline alone makes the claim implausible: Fable only became publicly available on July 1st, making it virtually impossible to distill enough data, train a model, and release it within two weeks. Nathan Lambert of the Allen Institute for AI added that distillation's impact has been declining as Chinese models approach the frontier and training increasingly relies on reinforcement learning — meaning that even if distillation data were used, it would not explain Kimi K3's capabilities. Lambert noted that supervised fine-tuning, where distillation artifacts like a model identifying as 'Claude' can appear, primarily shapes a model's 'manners,' not its core reasoning abilities.