The DeepSeek R1 reasoning model was released in January 2025, employs a rule-based reward reinforcement learning training strategy (GRPO algorithm), and is open-sourced under the MIT license