DeepSeek releases R1 reasoning models via reinforcement learning

DeepSeek introduces its first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1, which achieve performance comparable to OpenAI-o1-1217 on reasoning tasks. DeepSeek-R1-Zero is trained via large-scale reinforcement learning without supervised fine-tuning, while DeepSeek-R1 uses a multi-stage pipeline with cold-start data to improve readability.

• DeepSeek-R1-Zero naturally develops reasoning behaviors like self-verification and long chain-of-thought generation through pure RL.
• DeepSeek-R1 achieves 79.8% Pass@1 on AIME 2024, surpassing OpenAI-o1-1217, and scores 97.3% on MATH-500.
• The team open-sources six dense models (1.5B to 70B) distilled from DeepSeek-R1 based on Qwen and Llama architectures.
• Distilled models like DeepSeek-R1-Distill-Qwen-32B set new records among dense models, outperforming previous open-source competitors.

The release provides the research community with open-source checkpoints and an API to facilitate further development of smaller, powerful reasoning models.