DeepSeek-R1: An Open-Source LLM That Can Reason

DeepSeek-R1 is a cutting-edge large language model (LLM) that boasts impressive reasoning capabilities. Unlike typical LLMs that simply predict the next word, DeepSeek-R1 generates 'thinking tokens' to systematically solve problems. Its training involves three stages: first, a base model is trained on massive datasets; second, supervised fine-tuning using 600,000 long chain-of-thought reasoning examples generated by a specialized reasoning model; and finally, reinforcement learning to enhance both reasoning and non-reasoning task performance. DeepSeek-R1's success demonstrates that combining high-quality base models with automatically verifiable reasoning tasks significantly reduces reliance on labeled data, paving the way for future LLM advancements.