DeepSeek R1

DeepSeek-R1 is the first-generation reasoning model from DeepSeek, released January 20, 2025. It is a 671B-total / 37B-active MoE model trained from DeepSeek-V3-Base using large-scale reinforcement learning (RL) with Group Relative Policy Optimization (GRPO). Unlike conventional post-training that starts with supervised fine-tuning, R1-Zero was trained purely via RL without any SFT, allowing the model to discover its own chain-of-thought patterns. R1 refines this foundation with rejection sampling, additional RL stages, and SFT for human preference alignment.

For TheRouter operators, R1 is the canonical open-weight reasoning model. It delivers performance on par with OpenAI-o1 across math, coding, and complex reasoning benchmarks while remaining fully open-source under the MIT license. The model natively outputs <think>…</think> reasoning traces, enabling transparent step-by-step problem solving. A May 28, 2025 refresh (R1-0528) added JSON output, function calling, reduced hallucinations, and improved front-end capabilities. DeepSeek also released six distilled smaller models (1.5B–70B) that inherit R1 reasoning patterns.

The September 18, 2025 cover feature in Nature (Vol. 645, Issue 8081) marks the first time a mainstream large language model passed rigorous peer review at a top-tier scientific journal. The paper details how pure RL can incentivize emergent reasoning without human-annotated traces, a landmark result that challenges assumptions about training cost and transparency in frontier AI.