DeepSeek-R1 in Nature (peer-reviewed)
R1 becomes the first major LLM paper through peer review; supplement discloses roughly $294K RL cost on top of the base model.
Journal version adds reviewer-driven clarifications on training data, safety, and a denial of copying OpenAI reasoning traces. The ~$294K covers R1's RL stage only; the underlying base model cost about $6M separately, a distinction often lost in headlines.
- Date
- Wednesday, 17 September 2025
- Lab
- DeepSeek
- Kind
- paper
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| R1 RL training cost | ~$294,000 RL stage only; base model ~$6M separate; trained mainly on H800s | company |
Nature vol. 645, pp. 633-638 (2025). Nature's own page blocks automated fetching; date and details taken from Scientific American's report and the arXiv record's journal reference. Cost figure is company-reported.
Sources
- www.scientificamerican.com/article/secrets-of-chinese-ai-model-deepseek-revealed-in-landma
- arxiv.org/abs/2501.12948
This record was checked against its sources on 6 October 2026. How we check