"Kimi and DeepSeek distilled GPT-6 Astra," The Information reports; replies ask "who's lecturing whom?"

Moonshot AI's Kimi K3.1 Pro and DeepSeek's V4.1-Flash 1017 were trained on outputs from OpenAI's GPT-6 Astra, The Information reported on Oct. 28. Reuters relayed the report the same day, citing people with direct knowledge. Distillation means using an existing model's answers as training data for another model. Neither Moonshot nor DeepSeek commented right after the report.

It isn't the first such allegation. In a Sept. 30 blog post, OpenAI linked a key part of a distillation campaign that began in July to individuals associated with Moonshot AI. Anthropic said earlier this month that new protections cut data harvested through distillation-style attacks by 90% to 99%. A post by AI news account Chubby sharing the report drew 1,100 likes.

Replies split. One user wrote that OpenAI and Anthropic trained on nearly all of humanity's data without permission and then get angry when someone distills them; another said it was awkward for an industry that scraped half the internet to be outraged by distillation (150 likes).

There were technical takes too. Tiny corp wrote that "Astra produces reasoning in a clumsy mixed English, and I see that in Kimi K3.1 too," but added that "a bit too much RL doesn't necessarily mean distillation." Finimize noted that if labs lock down model access, legitimate high-volume API customers could face higher costs and more friction.