Moonshot AI and DeepSeek reportedly used outputs from OpenAI’s GPT-6 Astra to train the new open-weight models they launched in October 2026. DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17 with sharper task performance and improved agentic coding, while Moonshot AI launched Kimi K3.1 Pro on October 23, a model with approximately 2.8 trillion parameters. Kimi K3.1 Pro scored 55 on the Artificial Analysis Intelligence Index, ranking fourth overall, and outperformed GPT-6 Astra on several agentic benchmarks, including Terminal-Bench 4.0. The practice, known as distillation, previously drew accusations from Anthropic in February 2026 that the two labs ran distillation attacks against Claude and used fraudulent accounts to extract its outputs. Moonshot AI’s reported valuation of about $50 billion ahead of a planned Hong Kong IPO and DeepSeek’s higher valuation make training practices a sensitive issue for investors.
Source: Read the original article


