The DeepSeek-R1 series presents groundbreaking reasoning models that push the boundaries of performance in math, code, and reasoning tasks. DeepSeek-R1-Zero utilizes reinforcement learning (RL) to achieve remarkable results, although it faces challenges like endless repetition. DeepSeek-R1 improves upon this by incorporating cold-start data before RL, outperforming OpenAI-o1-mini. The models are available for open-source use, including distilled versions with impressive results compared to larger models. The breakthrough in incentivizing reasoning capabilities purely through RL sets a precedent for future advancements. The DeepSeek-R1 series is a game-changer in the research community, offering new possibilities for model development.
https://github.com/deepseek-ai/DeepSeek-R1