DeepSeek-R1-0528
R1 upgrade with deeper reasoning (about 23K tokens per AIME problem vs 12K) that lifts AIME 2025 from 70% to 87.5%, and adds system prompts and function calling.
More RL compute on the same V3 base roughly doubled reasoning length, lifting competition math and code scores; fewer hallucinations, JSON output, function calling and system-prompt support added. MIT license.
- Date
- Wednesday, 28 May 2025
- Lab
- DeepSeek
- Kind
- open-weights
- Access
- open weights
- Price
- unchanged vs R1 on the API
Figures
| Measure | Value | Measured by |
|---|---|---|
| AIME 2025 | 87.5% (from 70%) HMMT 2025 41.7 to 79.4; LiveCodeBench 63.5 to 73.3 | company |
Sources
This record was checked against its sources on 6 October 2026. How we check