DeepSeek and Moonshot Reportedly Trained on OpenAI's Astra

DeepSeek and Moonshot AI relied heavily on outputs from OpenAI’s GPT-6 Astra to train recent models, according to a report citing people familiar with the matter. The reported use, which has not been independently verified, concerns DeepSeek’s V4.1Flash 1017 and Moonshot’s Kimi K3.1 Pro, both released in October. The process, known as distillation, involves prompting a more capable model, collecting its responses and using them to train another system. Developers can use the method to reproduce some capabilities without building them entirely from scratch. The report did not establish how much Astra output was used or independently document the training process. DeepSeek launched V4.1Flash 1017 on October 17. Moonshot released Kimi K3.1 Pro on October 23. Neither company confirmed the reported use in the accounts provided. Moonshot declined to comment in one account, while DeepSeek did not respond to a request for comment; another account said both companies did not immediately respond. OpenAI has previously described a separate campaign that it said began in July and was linked in part to people associated with Moonshot. In a September 30 statement, OpenAI said the activity...