DeepSeek-R1

DeepSeek’s reasoning model trained with reinforcement learning to produce long, often visible chain-of-thought traces before the final answer.

How it works

DeepSeek-R1 is trained with Group Relative Policy Optimisation (GRPO), a reinforcement learning approach that rewards correct final answers on verifiable tasks like maths and coding. The model learns to generate extended chain-of-thought reasoning traces before committing to an answer. Crucially, reasoning ability emerged primarily from the RL process rather than from supervised imitation of human reasoning — a finding with significant implications for how reasoning can be instilled in language models.

Why it matters

DeepSeek-R1 matched or exceeded OpenAI's o1 reasoning model on many benchmarks while being fully open-weight and far cheaper to run. This showed that high-quality reasoning models are not restricted to the largest closed labs and significantly accelerated open-source reasoning research. Its distilled variants made strong reasoning capabilities accessible on consumer hardware for the first time.