DeepSeek, Moonshot used OpenAI’s GPT-6 Astra to train latest models: Report

Chinese AI developers DeepSeek and Moonshot AI have adopted a technique called distillation, which involves utilizing outputs from OpenAI's powerful GPT-6 Astra model to train their latest models, according to a report by The Information. This practice has emerged as a contentious issue in the ongoing US-China technology rivalry, as companies argue that distillation enables rivals to replicate advanced AI capabilities more swiftly and at a lower cost.

Moonshot AI's recently launched Kimi K3.1 Pro model and DeepSeek's upgraded V4.1-Flash 1017 model both utilized distillation from Astra. OpenAI, DeepSeek, and Moonshot AI have yet to comment on the matter. In a September 30 blog post, OpenAI disclosed that it had linked a core component of its distillation campaign, initiated in July, to individuals affiliated with Moonshot AI.

Major players in the AI industry, such as Anthropic, OpenAI, and Google, contend that distillation allows competitors to swiftly and cost-effectively replicate cutting-edge AI capabilities. Anthropic recently unveiled new measures aimed at thwarting distillation attacks, which they claim reduced data harvested from such attacks by 90% to 99% earlier this month.