Moonshot Accused of Distilling Fable to Build Kimi K3
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Michael Kratsios, White House science advisor, accused Moonshot of building Kimi K3 by copying Anthropic's Fable LLM and using Nvidia Grace Blackwell 300 chips not cleared for export to China, calling the alleged distillation "unacceptable."
- Treasury Secretary Scott Bessent separately claimed "watermarks of our U.S. large language models" were found on Chinese models, though he did not specify what those watermarks consist of and the Treasury Department did not respond to queries.
- Braden Hancock (Laude Institute) and Nathan Lambert (Allen Institute for AI) said the two-week gap between Fable's July 1 public release and Kimi K3's arrival makes distillation technically implausible — there wasn't enough time to query, train, and release a frontier model.
- Lambert said SFT-based distillation benefits are diminishing as models grow more complex, and replicating frontier capabilities would require reinforcement learning runs that would be "insanely expensive" and potentially time-prohibitive via API access.
- Anthropic publicly accused Moonshot, DeepSeek, and MiniMax of systematic distillation earlier this year, claiming it identified millions of exchanges from those companies via IP addresses and metadata "distinct from normal usage patterns."
- Distillation is an industry-wide practice, not limited to China — Elon Musk testified that SpaceXAI distilled OpenAI models to build Grok, and Hancock warned that "Americans are understating the technical expertise of these Chinese teams."
- Sam Bresnick (Georgetown's Center for Security and Emerging Technology) said a black market exists for restricted Nvidia chips, noting that Supermicro's founder was indicted in May for smuggling advanced chips into China; Biden's Commerce Department proposed know-your-customer rules for data centers in 2024 but no progress has been made under Trump.
Why it matters: The US case for potentially banning Chinese open-weight models rests on distillation allegations that experts publicly call technically implausible given Kimi K3's two-week development window. If the central evidence doesn't withstand scrutiny, it weakens the regulatory rationale while the parallel policy track — KYC reporting rules for data centers handling restricted chips — remains stalled, leaving enforcement reliant on chip export controls that already have a documented smuggling problem.
