Experts doubt distillation alone built Kimi K3, despite U.S. official claims

White House science advisor Michael Kratsios leveled serious accusations at Moonshot, the Chinese company behind the Kimi K3 LLM, claiming it reverse-engineered Anthropic's Fable model using smuggled semiconductor chips. Yet leading AI researchers are questioning whether distillation—the technique of querying a model to extract and replicate its capabilities—could realistically account for K3's performance in such a compressed timeframe.

The official claim#

Kratsios characterized the alleged activity as "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research." Treasury Secretary Scott Bessent reinforced the narrative, noting that watermarks from U.S. language models have been detected in Chinese competitors. Moonshot did not respond to inquiries about its training methodology, and neither official provided additional corroborating evidence.

Why experts remain skeptical#

Braden Hancock, a researcher at the Laude Institute and Snorkel AI co-founder, points to a fundamental problem: Fable only entered public availability on July 1st, yet K3 emerged roughly two weeks later. "You can't distill that much data, train a model, and release it in two weeks," Hancock told TechCrunch. Building a frontier-class model through distillation alone would demand weeks or months of sustained work, not days.

Nathan Lambert, an AI researcher at the Allen Institute for AI, adds that distillation's utility diminishes as Chinese models approach the performance frontier. Contemporary systems increasingly rely on reinforcement learning rather than simple fine-tuning to extract capabilities. That process demands vastly more compute and could prove prohibitively expensive or slow when channeled through an external API.

The evidence gap#

Anthropic itself has documented distillation attempts against its models and publicly named Moonshot as a target earlier this year, identifying millions of deliberate queries from suspected company addresses. However, no direct technical analysis linking K3's architecture or performance gains to Fable has been shared publicly.

Distillation remains widespread across the industry, not confined to Chinese labs. Elon Musk testified earlier this year that his company distilled OpenAI models to build Grok, framing the practice as routine in AI development.

The smuggled chips angle#