A public statement from Michael Kratsios alleges that Moonshot AI covertly distilled Anthropic’s Fable model to build its own K3 system. The accusation claims the developer ran a sophisticated internal platform to extract data from American models while bypassing detection systems. But Moonshot AI has yet to respond to these claims of intellectual property theft.
We have information that Moonshot AI distilled Anthropic’s Fable for the development of its K3 model.
— Director Michael Kratsios (@mkratsios47) July 22, 2026
To do this they developed a sophisticated internal platform to conduct large scale distillation against U.S. models, allowing them to quickly switch between multiple methods of…
The allegations outline a highly organized operation designed to siphon proprietary technology. According to the post by Kratsios on the social media platform X, the Chinese firm built an evasion system to rapidly switch between different access points. This allowed them to scrape data from US models without triggering rate limits or security flags. While distillation is common for making smaller models, the scale of this operation has drawn sharp criticism from US officials.
Apart from software scraping, the firm is also accused of bypassing hardware restrictions. The report indicates that Moonshot AI acquired servers containing GB300 chips and accessed identical processors hosted in Thailand. These chips are subject to strict export controls. Accessing them in other countries allows developers to train heavy models outside the reach of domestic regulations. It highlights a growing trend of companies finding creative ways to secure computational power.
Federal policy supports open research and fair competition in the AI sector. Legitimate model optimization helps build efficient consumer products. However, covert industrial espionage is viewed as a threat to national research. US officials maintain that open source development must coexist with strong protection for proprietary breakthroughs.
