DeepSeek V3.1
llmA hybrid model merging DeepSeek V3 and R1 into a single checkpoint with a switchable thinking and non-thinking chat template.
Metadata
| org | DeepSeek |
| released | 2025-08-21 |
| params | 671B-A37B |
| license | mit |
| kind | llm |
| hf id | deepseek-ai/DeepSeek-V3.1 |
Benchmarks
| GPQA | 80.1 | source ↗ |
| AIME 2025 | 88.4 | source ↗ |
| SWE-bench | 66 | source ↗ |
Lineage
↑ finetuneDeepSeek V3
↓ finetuneDeepSeek V3.2-Exp