DeepSeek-V2.5
Merged DeepSeek-V2-Chat and Coder-V2 into one model served on both API endpoints.
One model for chat and code, with a safety score of 82.6% up from 74.4% and 5.1% better fill-in-the-middle completion. A consolidation step toward the single-model API that V3 formalised.
- Date
- Thursday, 5 September 2024
- Lab
- DeepSeek
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| Safety score | 82.6% Up from 74.4%; FIM completion +5.1% | company |
V2.5-1210 (2024-12-10) was the final V2.5 refresh, with better math and coding and web search in the app.
Sources
This record was checked against its sources on 6 October 2026. How we check