DeepSeek V4 Pro (Non-reasoning): benchmark results
DeepSeek V4 Pro non-reasoning evaluation mode. Provider: DeepSeek. Released 2026-04-24. Access: Open.
Aggregate Rating 1656 ± 11, rank #504 of 2168 rated models, from 96 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| NetConfArena - Configuration F1 | 79.6 | Configuration line F1 against reference (%) | 100 |
| NetConfArena - Task Pass Rate | 85.2 | Instances with every test case satisfied (%) | 100 |
| NetConfArena - Test-Case Score | 96.1 | Mean fraction of task test cases satisfied (%) | 100 |
| PIJ - Criminal Profiling | 51.38 | F1 (%, 50-sample human evaluation, include Unknown) | 100 |
| TRACE Bench - Diversity | 95.43 | Normalized repetition-penalty score (%) | 100 |
| AA Omniscience - Software Engineering (SWE) - Java | 36 | Accuracy (%) | 95.7 |
| AA Omniscience - Software Engineering (SWE) - PHP | 50 | Accuracy (%) | 93.9 |
| DocFormBench (DocFormFlow) - Hallucinated Formatting | 1.57 | Hallucinated formatting rate (%) on DocFormBench: share of a | 92.9 |
| VitaBench 2.0 (RAG Memory) | 42.4 | Avg@4 (%, times 100): mean task success over four runs at te | 92.3 |
| AA Omniscience - Software Engineering (SWE) - TypeScript | 47.78 | Accuracy (%) | 92 |
| UGI - Natural Intelligence | 59.21 | NatInt Score | 91.9 |
| PredAct-Bench | 60.7 | Final F1 (%, PredAct-CS) | 91.7 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-pro-non-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-20.