White House calls Chinese AI lab a copycat, but experts say Kimi K3 can't just be a Claude clone

The White House is lobbing a serious accusation at one of China’s most prominent AI labs. Science advisor Michael Kratsios claims Moonshot, the company behind the massive open-weight model Kimi K3, built its creation by illicitly copying Anthropic’s Fable LLM using smuggled Nvidia chips. “Large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology and undermining American research is unacceptable,” Kratsios wrote, echoing Treasury Secretary Scott Bessent’s claim that they’re “finding watermarks” of American models on Chinese ones. Moonshot hasn’t commented, and the government hasn’t provided any evidence for the watermark claim.

But the technical community isn’t buying the simple copycat narrative. Researchers point to a glaring timeline problem. Anthropic’s Fable only became publicly available on July 1st, and Kimi K3 appeared just two weeks later. “You can’t distill that much data, train a model, and release it in two weeks,” Braden Hancock of the Laude Institute and Snorkel AI told TechCrunch. The sheer logistical impossibility of that schedule is making experts skeptical that distillation alone explains the model’s advanced capabilities.

The debate hinges on how modern AI models actually learn. Distillation is a known technique where a lab systematically queries a target model to extract its problem-solving methods, often through supervised fine-tuning (SFT). But Nathan Lambert, a researcher at the Allen Institute for AI, argues that SFT’s importance is fading as models reach the frontier and shift to reinforcement learning. To truly copy Fable-level reasoning, a lab would need to run millions of expensive, slow API queries through a reinforcement learning agent—a process that would be both a financial and time bottleneck. “If it were the case, everyone would be easily able to catch up,” Lambert said on a podcast, noting we haven’t seen that happen.

The reality is probably messier than a simple heist. Anthropic has previously accused Moonshot, DeepSeek, and MiniMax of systematic distillation earlier this year, discovering millions of unusual queries. But distillation is an industry-wide gray area, not a uniquely Chinese sin. Elon Musk testified that his own team distilled OpenAI models to build Grok, calling it common practice. Hancock adds that American observers consistently underrate Chinese engineering talent, noting Moonshot’s founder is a CMU PhD. “If American models ground to a halt, I think China’s progress would slow, but would still continue,” he said. “They’re not just riding coattails here.” The chip-smuggling allegation adds another layer, but it doesn’t settle the core question of whether Kimi K3 is a copycat or evidence of genuine, independent technical momentum.