DeepSeek V4 Flash: benchmark results
DeepSeek's efficient open V4 Flash MoE model (13B active). Provider: DeepSeek. Released 2026-04-24. Access: Open.
Aggregate Rating 1734 ± 3, rank #137 of 1656 rated models, from 1647 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AARRI-Bench (Hermes Agent) - Context | 55.9 | Task success rate (%) under the 0/1 reward on the context ta | 100 |
| AARRI-Bench (Hermes Agent) - Interaction | 76.2 | Task success rate (%) under the 0/1 reward on the interactio | 100 |
| ANIMA (AnimalHarmBench 2.0) | 74.4 | Overall score (0-100; multi-judge LLM grading of reasoning a | 100 |
| ATRBench | 23.7 | TSAcc Default (%) (self-reported) | 100 |
| AutoRO-Bench (AutoREM) | 97.4 | Accuracy (%; share of the 64 in-distribution Random robust-o | 100 |
| BioPhys-Bridge | 0.36 | Evidence-ID F1 (0-1) | 100 |
| CAP (Cross-Site Browser Tasks) | 25 | Partial completion (%; mean root-node score of the rubric tr | 100 |
| CAP (Cross-Site Browser Tasks) - Complex Actions | 33.7 | Action score (%; mean leaf score over the execution checkpoi | 100 |
| CAP (Cross-Site Browser Tasks) - Perception | 34.2 | Perception score (%; mean leaf score over the perception che | 100 |
| CLIR-Bench - Threshold Forecasting | 38.33 | Accuracy (%) on threshold forecasting (will a variable cross | 100 |
| CUE-Bench - Affective Stance | 49.2 | Accuracy (%; nine-way affective stance recognition over expl | 100 |
| CUE-Bench - Pragmatic Intent | 32.5 | Accuracy (%; pragmatic intent classification; over CUE-Bench | 100 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-flash · How It Works · Data refreshed daily, snapshot 2026-10-20.