
Chinese AI developers DeepSeek and Moonshot AI relied heavily on outputs from OpenAI's GPT-6 Astra model to train their latest models, according to a report by The Information.
The practice of model distillation has become a key flashpoint in US-China technology competition as leading firms argue it allows rivals to reproduce advanced capabilities quickly and cheaply.
Moonshot AI's Kimi K3.1 Pro and DeepSeek's upgraded V4.1-Flash 1017 model both benefited from distilling Astra, while OpenAI previously linked distillation efforts to individuals associated with Moonshot.