DeepSeek-R1

DeepSeek-R1 is the first openly-released reasoning model that was competitive with OpenAI o1, and the paper that accompanied it (published Jan 22, 2025) became one of the most-read AI papers of the year. The technical story was that pure reinforcement learning post-training, without any supervised fine-tuning on human reasoning traces, could elicit chain-of-thought reasoning that generalized to held-out problems. The R1-Zero variant showed this most cleanly: starting from the V3 base, the team applied Group Relative Policy Optimization (GRPO) with a verifiable reward function and watched the model spontaneously develop longer reasoning traces over training. The full R1 added a small SFT cold-start to clean up readability before the RL phase. MIT-licensed weights and several distillations into smaller dense bases (1.5B, 7B, 8B, 14B, 32B, 70B Llama and Qwen variants) followed in the same release. The R1 distillations into 32B and below put genuinely capable reasoning models within reach of local deployment for the first time, and the GRPO recipe became the template that Qwen 3, Llama 4, and several Western labs followed in their own reasoning post-training.