Key facts
- Moonshot AI's Kimi K3.1 Pro model and DeepSeek's V4.1-Flash 1017 model benefited from distilling Astra.
- Distillation allows rivals to reproduce advanced capabilities more quickly and cheaply.
- Anthropic reported new measures reduced data harvested in distillation attacks by 90% to 99%.
Chinese AI developers DeepSeek and Moonshot AI relied on outputs from OpenAI's GPT-6 Astra model to train their latest AI models, according to a report by The Information citing people with direct knowledge. This practice, known as distillation, allows companies to replicate advanced AI capabilities more rapidly and at a lower cost. Moonshot AI's Kimi K3.1 Pro, released last week, and DeepSeek's upgraded V4.1-Flash 1017 model, launched on October 17, reportedly benefited from this method. OpenAI had previously indicated in a September blog post that it had linked a distillation campaign to individuals associated with Moonshot AI. Other firms like Anthropic and Google have also voiced concerns about this practice, with Anthropic reporting significant reductions in data harvesting through new preventative measures.