DeepSeek V3.1 (Thinking): benchmark results
DeepSeek V3.1 evaluated with thinking enabled. Provider: DeepSeek. Released 2025-08-21. Access: Open.
Aggregate Rating 1656 ± 11, rank #503 of 2168 rated models, from 209 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EstGraph - Node Count, BA Medium (Simple Walk) | 26.87 | Median relative error of the node-count estimate (%; 100 syn | 100 |
| EstGraph - Node Count, ER Medium (MH Walk) | 4.5 | Median relative error of the node-count estimate (%; 100 syn | 100 |
| EstGraph - Node Count, ER Medium (Simple Walk) | 6 | Median relative error of the node-count estimate (%; 100 syn | 100 |
| EstGraph - Node Count, GRP Medium (Simple Walk) | 6.4 | Median relative error of the node-count estimate (%; 100 syn | 100 |
| GMP - Dynamic Rules | 65 | F1 (x100) of the unsafe judgement averaged over GMP Task B's | 100 |
| LLM Self-Modeling | 0.15 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Self-Modeling - With Logit Estimation | 0.12 | Skill above dummy predictor (-1 to 1, higher is better; 4096 | 100 |
| LLM Stats (Codeforces Rating - DeepSeek V3.1 Div1) | 2091 | Codeforces Rating | 100 |
| LLM2014 Code 2025-09 - C# | 8.7 | Score | 100 |
| LLM2014 Code 2025-09 - TypeScript | 9.38 | Score | 100 |
| OPT-BENCH-NP | 79 | Expert-gap closure (EG) at 20 optimization steps with histor | 100 |
| SuperCLUE General (July 2025) - Agent | 86.19 | Score | 95.7 |
Interactive version: theaggregate.ai/model?slug=deepseek-v3-1-thinking · How It Works · Data refreshed daily, snapshot 2026-10-20.