DeepSeek releases open reasoning model R1

Chinese artificial-intelligence company DeepSeek released its R1 reasoning model on 20 January 2025, publishing downloadable model weights and a technical paper. The system was designed to spend additional computation on multi-step problems and showed strong results on several mathematics, coding and reasoning benchmarks.

DeepSeek described a training approach using reinforcement learning and reported development costs substantially below common assumptions for frontier models. Cost comparisons require care because disclosed figures may cover a training run rather than all research, data, hardware and experimentation.

The model's license allowed broad use of weights, while the training data and full infrastructure were not completely open. “Open weights” is therefore more precise than claiming the entire system was open source. R1 could still generate false information, biased content and unsafe outputs.

The release drew global attention to Chinese AI capability and affected technology-company valuations. Market movement is evidence of investor reaction, not proof that one model made all existing infrastructure obsolete. Performance also varies by task, prompt and evaluation quality.

DeepSeek's repository and paper establish availability and reported methods. This item records a model release, distinct from the company's earlier app and later revisions. It explains the competitive significance without accepting unverified cost claims or treating benchmark scores as general intelligence.