DeepSeek-R1 published in Nature
R1 paper published in Nature after peer review, a first for a major LLM; discloses a $294K RL cost on top of the V3 base.
Peer-reviewed version adds supplementary detail on training cost, data and safety. DeepSeek states R1 did not learn by copying OpenAI reasoning outputs; reviewers found the rebuttal credible. $294K covers RL runs only, not the V3 base (about $5.6M).
- Date
- Wednesday, 17 September 2025
- Lab
- DeepSeek
- Kind
- paper
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Reported R1 RL training cost | $294K 512 H800s ran 198 hours for R1-Zero plus about 80 for R1. This excludes the V3 base (~$5.58M) | company |
Nature volume 645, pages 633-638 per the arXiv record; I could not open nature.com directly (cookie redirect), so numbers rely on Scientific American and The Register reporting of the paper.
Sources
- www.scientificamerican.com/article/secrets-of-chinese-ai-model-deepseek-revealed-in-landma
- www.theregister.com/2025/09/19/deepseek_cost_train/
- arxiv.org/abs/2501.12948
This record was checked against its sources on 6 October 2026. How we check