DeepSeek releases R1 reasoning model with open weights
Hangzhou-based DeepSeek published R1 and six distilled variants under the MIT license, claiming performance on par with OpenAI's o1 on math, code and reasoning.
The weights are public, the license permits commercial use and distillation, and the accompanying paper describes the reinforcement-learning recipe in enough detail for others to study it. The benchmark comparisons with o1 are DeepSeek’s own.
Index treatment. No milestone moves. The SI Index tracks the frontier, not how widely it is shared, and R1 claimed parity with a model already counted at the epoch.