The Information: DeepSeek and Moonshot AI used GPT-6 Astra to train their latest models
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
According to The Information, Chinese labs used the answers of an OpenAI model to train their own; DeepSeek and Moonshot do not comment on it.
According to The Information, citing people familiar with the matter, as The Decoder relays, the Chinese labs DeepSeek and Moonshot AI relied on OpenAI's GPT-6 Astra when training their latest models; Kimi K3.1 Pro and V4.1-Flash 1017 are named. The method, distillation, requires a model to reveal its hidden reasoning steps, and the terms of use of OpenAI and Anthropic prohibit this. Moonshot declined to comment, and DeepSeek did not respond to requests. According to The Decoder, OpenAI reported a "coordinated distillation campaign" in late September involving people linked to Moonshot. It adds that after Anthropic strengthened its defenses (new methods, by these accounts, cut data extraction by 90–99%), Astra became the main target; an OpenAI spokesperson said the company "takes the problem very seriously."
Sources 1