We value your privacy
This site uses cookies to improve your browsing experience, analyze site traffic, and show personalized content.
We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrated remarkable performance on reasoning. With RL, DeepSeek-R1-Zero naturally emerged with numerous powerful and interesting reasoning behaviors. However, DeepSeek-R1-Zero encounters challenges such as endless repetition, poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1 across math, code, and reasoning tasks. To support the research community, we have open-sourced DeepSeek-R1-Zero, DeepSeek-R1, and six dense models distilled from DeepSeek-R1 based on Llama and Qwen. DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI-o1-mini across various benchmarks, achieving new state-of-the-art results for dense models.
| $20.12% | |
| $20.08% | |
| $10.07% | |
| $10.06% | |
| $0.700.04% | |
| $0.630.03% | |
| $0.570.03% | |
| $0.520.03% | |
| $0.490.02% | |
| $0.400.02% | |
| $0.400.02% | |
| $0.350.02% |