DeepSeek and Moonshot reportedly trained new models on GPT-6 Astra outputs

DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train their newest models, The Information reported, citing people with direct knowledge. Reuters relayed the report and said the three companies did not immediately respond to requests for comment.

The models named in the report are Moonshot’s Kimi K3.1 Pro and DeepSeek’s upgraded V4.1-Flash 1017. The Information says both benefited from distillation: using outputs from a more capable model to improve another system more quickly and cheaply.

The claim lands amid a widening fight over who can copy frontier-model capabilities from API or chat outputs. OpenAI previously said it disrupted a coordinated campaign intended to extract protected reasoning and attributed a core cluster of that activity to people associated with Moonshot. It described the activity as adversarial distillation and said operators used coordinated model interactions rather than breaching databases or stored user conversations.

Model makers argue that large-scale unauthorized distillation can transfer advanced capabilities without the original investment in training and safety. If the report is accurate, it also complicates comparisons of inexpensive Chinese models with US frontier systems: some of the apparent efficiency may come from learning directly from a rival’s outputs rather than from an entirely independent training path.