The team introduces its first - generation reasoning models, DeepSeek - R1 - Zero and DeepSeek - R1. DeepSeek - R1 - Zero, trained via large - scale RL without SFT, shows remarkable reasoning ability but also has problems like endless repetition. DeepSeek - R1, which incorporates cold - start data before RL, solves these issues and achieves performance on par with OpenAI - o1. The team has open - sourced these models and six distilled ones, with DeepSeek - R1 - Distill - Qwen - 32B outperforming OpenAI - o1 - mini in benchmarks.
