DeepSeek-R1 is a groundbreaking first-generation reasoning model that outperforms OpenAI-o1 across math, code, and reasoning tasks. DeepSeek-R1-Zero, trained via reinforcement learning (RL) without supervised fine-tuning, showcased impressive reasoning capabilities, although it faced challenges like poor readability. To address these issues, DeepSeek-R1 incorporates cold-start data before RL training, achieving performance comparable to top AI models. The model’s ability to distill reasoning patterns from larger models into smaller ones demonstrates its versatility and potential impact on the research community. DeepSeek-R1 models are open-sourced, supporting commercial use and modifications. Additionally, users are provided with usage recommendations for optimal performance.
https://github.com/deepseek-ai/DeepSeek-R1