DeepSeek-R1

DeepSeek-R1 is an open reasoning model released by DeepSeek-AI on January 20, 2025. Architecturally it is a Mixture-of-Experts with 671 billion total parameters and 37B active per token, built on top of DeepSeek-V3-Base. R1 was produced via a Reasoning RL pipeline with verifiable rewards — GRPO (Group Relative Policy Optimization), where rewards come from rules (math correctness, code execution, format) rather than a learned reward model. Context: 128,000 tokens. Licence: MIT on the weights, the model is publicly available on Hugging Face.

Released alongside R1 was R1-Zero — a variant trained with pure RL without any SFT cold-start, proving that Reasoning RL can elicit long chain-of-thought and self-correction directly from RL. Production R1 added a short SFT cold-start on a few hundred CoT examples for more readable outputs. A series of distilled variants on Llama 3.1 and Qwen 2.5 (1.5B, 7B, 8B, 14B, 32B, 70B) was also published, porting much of R1's capability to single-GPU models.

Results: AIME 2024 79.8% pass@1, MATH-500 97.3%, Codeforces 96.3 percentile, MMLU 90.8%, GPQA Diamond 71.5%, LiveCodeBench 65.9%, SWE-bench Verified 49.2% — on par with or above OpenAI o1 at a fraction of inference cost. The model is available via the DeepSeek API, Hugging Face, Together AI, Fireworks, OpenRouter, Amazon Bedrock Marketplace and Vertex AI Model Garden. DeepSeek-R1 — together with the GRPO algorithm publication — established the de-facto standard for open Reasoning RL and triggered a wave of reproductions (TinyZero, Open-R1, SimpleRL).