DeepSeek announced R1-Lite-Preview, the first widely noticed non-OpenAI reasoning model with visible thoughts
DeepSeek announced R1-Lite-Preview on its chat site, advertising a "transparent thought process in real-time," o1-preview-level scores on AIME and MATH, and a chart of AIME accuracy rising as thought…
- Date
- 20 November 2024
- Who
- DeepSeek
- Confidence
- High
- 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 · Confidence: High DeepSeek announced R1-Lite-Preview on its chat site, advertising a "transparent thought process in real-time," o1-preview-level scores on AIME and MATH, and a chart of AIME accuracy rising as thought length grows; open-source models and an API were promised as coming soon (DeepSeek announcement, DeepSeek API docs; neither page mentions the free-tier message cap or a toggle name that some coverage reported, so those details are omitted here). It arrived 69 days after o1-preview and, unlike OpenAI, displayed the raw chain of thought, which is what drew attention. "First" here means first widely noticed from a major lab, because smaller open projects that emitted long o1-style thoughts came earlier (Open-O1, 2024-10-05; GAIR's "O1 Replication Journey," 2024-10-08 (arXiv 2410.18982); Alibaba's Marco-o1 followed a day later, 2024-11-21). The later V3 report confirms an internal "R1 series" model existed by then, since DeepSeek-V3's post-training distilled reasoning from it (distillation means training one model on another's outputs; V3 report §5.4.1). The full weights followed 61 days later as R1.