DeepSeek V4 Flash (Reasoning, High Effort): benchmark results
DeepSeek V4 Flash reasoning high-effort evaluation mode. Provider: DeepSeek. Released 2026-04-23. Access: Open.
Unified ELO 1639 ± 1, rank #285 of 1761 rated models, from 34 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA TAU-2 Bench | 95.61 | Accuracy (%) | 95.3 |
| AA Omniscience - Software Engineering (SWE) - PHP | 52 | Accuracy (%) | 94.9 |
| AA Omniscience - Software Engineering (SWE) - TypeScript | 50 | Accuracy (%) | 94.2 |
| AA Omniscience - Software Engineering (SWE) - Python | 41 | Accuracy (%) | 92.2 |
| AA IFBench | 73.47 | Accuracy (%) | 91.1 |
| AA Omniscience - Software Engineering (SWE) - R | 34 | Accuracy (%) | 91 |
| AA Omniscience - Software Engineering (SWE) - C | 60 | Accuracy (%) | 90.8 |
| AA Omniscience - Software Engineering (SWE) - Kotlin | 38 | Accuracy (%) | 90.4 |
| AA Omniscience - Software Engineering (SWE) - Julia | 32 | Accuracy (%) | 90.1 |
| AA Omniscience - Software Engineering (SWE) - JavaScript | 46.36 | Accuracy (%) | 88.4 |
| AA Omniscience - Software Engineering (SWE) - Dart | 32 | Accuracy (%) | 87.4 |
| AA Omniscience - Software Engineering (SWE) - Swift | 52 | Accuracy (%) | 86.7 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-flash-reasoning-high-effort · How It Works · Data refreshed daily, snapshot 2026-09-05.