Experts Question Claim That Kimi K3 Came From Anthropic Distillation
White House science adviser Michael Kratsios accused Moonshot AI of building Kimi K3 by distilling Anthropic's Fable model and using restricted Nvidia servers, the BBC reported. AI researchers have questioned whether distillation from Fable could explain the performance of Kimi K3, an open weight model that Moonshot unveiled last week.
Distillation means querying a stronger model to generate data for another model. Researchers cited the short period since Fable became publicly available on July 1 as a reason to doubt that Moonshot could query it at scale, train a new model, and release Kimi K3 within about two weeks.
The technical dispute centers on whether supervised fine tuning can reproduce advanced model capabilities. Researchers said that fine tuning can copy a model's style, but stronger capabilities would likely require reinforcement learning runs with large infrastructure and many model interactions.
The allegations also cover access to Nvidia GB300 Grace Blackwell servers, which are restricted for export to China. Moonshot has not publicly detailed its Kimi K3 training process.
We hope you enjoyed this article.
Consider subscribing to one of our newsletters like AI Policy Brief or Daily AI Brief.
Also, consider following us on social media:
More from: Regulation
Subscribe to AI Policy Brief
Weekly report on AI regulations, safety standards, government policies, and compliance requirements worldwide.
Market report
AI’s Time-to-Market Quagmire: Why Enterprises Struggle to Scale AI Innovation
The 2025 AI Governance Benchmark Report by ModelOp provides insights from 100 senior AI and data leaders across various industries, highlighting the challenges enterprises face in scaling AI initiatives. The report emphasizes the importance of AI governance and automation in overcoming fragmented systems and inconsistent practices, showcasing how early adoption correlates with faster deployment and stronger ROI.
Read more