DeepSeek-R1 is an open reasoning model from the AI lab DeepSeek that learned to think through problems step by step mostly by reinforcement learning.
In plain words
DeepSeek showed that, without human-written worked solutions, rewarding a model for correct answers on checkable tasks, like maths and code, can make long step-by-step reasoning appear by itself. DeepSeek-R1, released in January 2025, is a mixture-of-experts model with 671 billion parameters, 37 billion active per token. Its weights are under the MIT licence, and smaller distilled versions based on Llama and Qwen were released too.
A quick example
Give it a tricky maths problem and it first writes out a long chain of reasoning, checking and correcting itself, and only then gives the final answer.
Where you'll meet it
- Maths, coding and science questions with answers that can be checked.
- Research on reasoning and on distilling reasoning into small models.
Related
Sources
- PaperDeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, DeepSeek-AI (arXiv) · read 29 Sept 2026
- DocsDeepSeek-R1, DeepSeek (GitHub) · read 29 Sept 2026
Checked by WhatIsAI against the sources above · 29 Sept 2026. Spotted an error? Report it.