DeepSeek-R1 is published in Nature after peer review
R1 was published online in Nature on 2025-09-17 (print issue 2025-09-18, volume 645, issue 8081, pages 633-638; Crossref) after peer review.
- Date
- 17 September 2025
- Who
- DeepSeek, Nature
- People
- Liang Wenfeng and the DeepSeek-AI team; commentators named by Scientific American include Lewis Tunstall (Hugging Face) and Huan Sun (Ohio State)
- Confidence
- High on the publication; Medium on peer-review details (secondary reporting)
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): DeepSeek, Nature · People: Liang Wenfeng and the DeepSeek-AI team; commentators named by Scientific American include Lewis Tunstall (Hugging Face) and Huan Sun (Ohio State) · Confidence: High on the publication; Medium on peer-review details (secondary reporting) R1 was published online in Nature on 2025-09-17 (print issue 2025-09-18, volume 645, issue 8081, pages 633-638; Crossref) after peer review. Scientific American says it is "thought to be the first major LLM to undergo the peer-review process," which is a media framing and not a settled fact, since GPT-3 appeared at NeurIPS 2020 (paper) and PaLM in the Journal of Machine Learning Research in 2023 (JMLR), both peer-reviewed, so the claim that holds is the narrower one, first major frontier-class LLM reported to go through journal review. DeepSeek responded to reviewer feedback by reducing anthropomorphic language and clarifying technical details, and the supplementary materials disclosed a training cost of about $294,000 for the RL stage. DeepSeek said R1 did not learn by copying reasoning examples from OpenAI models, acknowledging that its web-crawled base data may contain AI-generated text; Huan Sun, who commented on the review process, told Scientific American the rebuttal was as convincing as anything seen in publications (Scientific American, which carries the reviewer, cost and distillation details; The Register also covered the paper but, as fetched, not those details; the Nature page itself redirects to a login hop; whether R1 was on the issue's cover is not verified here). The extended arXiv v2 (2026-01-04) carries the full supplement used in this chapter. What the figure does and does not cover is in B08-15. The Nature record also contrasts with closed labs' practice of publishing no method for their reasoning models (B08-01).