DeepSeek-R1
An open-weight reasoning model from DeepSeek, released in January 2025 under the MIT license, trained with large-scale reinforcement learning to think step by step before answering.
Also DeepSeek · R1 · reasoning model
DeepSeek-AI, 2025
A model that was rewarded for getting math and coding answers right, and on its own learned to write out long chains of reasoning, check itself and try again.
Where it fits
+All relationships as a list
Is a kind of / part of
Uses / built on
How it works
DeepSeek released R1 in January 2025, with weights under the permissive MIT license. It is a reasoning model: before answering, it writes out a long chain of thought, checks its work and corrects itself (see prompting for chain-of-thought).
What made it notable was how that behavior was trained. Starting from DeepSeek’s mixture-of-experts base model, the team ran large-scale reinforcement learning (policy-gradient style) on problems with checkable answers, such as math and code. In the experimental R1-Zero version, with no supervised examples of reasoning at all, long step-by-step reasoning emerged from the reward alone. R1 itself added a small supervised warm-up to make the reasoning more readable.
DeepSeek also released smaller models distilled from R1’s outputs, which brought reasoning-style behavior to models that run on ordinary hardware.
Everyday use cases
- Case 01
How you've probably met it
Through news coverage of its January 2025 release, the DeepSeek apps, or open models distilled from it that run locally.
Strengths, weaknesses and when to use it
Strengths
- Open weights for a strong reasoning model, permitting study, local use and distillation.
Weaknesses
- Long reasoning makes answers slower and more expensive to generate.
When to use itFor math, coding and multi-step reasoning, especially when you want to run or study the model yourself.
Compare with
Go deeper
Sources
Facts checked as of 2026-10-12