Deep Seek and Moonshot AI relied heavily on output from OpenAI's GPT-6 Astra model to train their newest releases, the Information reported on October 28, citing people with direct knowledge of the matter1,2.
Moonshot AI's Kimi K3.1 Pro, released the week before the report, and Deep Seek's upgraded V4.1-Flash-1017, launched on October 17, both used GPT-6 Astra output for distillation training. Distillation involves using one model's output to train another, enabling developers to reproduce advanced capabilities faster and at lower cost.
The report follows OpenAI's own September 30 blog post, in which the company said it had linked a core part of a distillation campaign that began in July to individuals associated with Moonshot AI. OpenAI, Deep Seek, and Moonshot AI did not immediately respond to Reuters' requests for comment on the Information's report.
The practice has become a point of contention among leading AI companies as US-China technology competition intensifies. Anthropic, OpenAI, and Google have argued that distillation allows rivals to replicate the capabilities of advanced models more quickly and cheaply. Anthropic said earlier in October that new measures designed to prevent rivals from copying the capabilities of its Claude models had reduced data harvested in distillation attacks by about 90 percent to 99 percent.
ANALYSIS The two distillation cases surfacing within weeks of each other put concrete examples behind the complaints US frontier labs have raised about the practice, and they arrive as OpenAI's own investigation had already publicly named Moonshot AI in connection with a July distillation campaign.