Live Updates > Live Update Details
2025-09-18 13:06:54
[DeepSeek responds to OpenAI's questions about distillation for the first time] On September 18, DeepSeek caused a sensation again. The DeepSeek-R1 research paper, jointly completed by the DeepSeek team and with Liang Wenfeng as the corresponding author, appeared on the cover of the internationally renowned journal Nature. In January of this year, DeepSeek published the first version of the preprint paper on arxiv. In comparison, the version published in Nature added more model details and reduced the anthropomorphic descriptions in the description. In the supplementary materials, DeepSeek mentioned that the training cost of the R1 model was only US$294,000, and responded to OpenAI's questions about distillation when the model was first released. In January of this year, it was reported that OpenAI researchers believed that DeepSeek may have used the output of the OpenAI model to train R1. This method can accelerate the improvement of model capabilities while using fewer resources. In the supplementary materials section of the paper, DeepSeek responded to questions about the source of DeepSeek-V3-Base training data. "The training data for DeepSeek-V3-Base comes only from ordinary web pages and e-books and does not contain any synthetic data. In the pre-training cooling phase, we did not intentionally add synthetic data generated by OpenAI. All data used in this phase was scraped from the web." DeepSeek said
