Qwen 3 VL 4B: benchmark results
Can locate interface elements for GUI agents, read text in images and generate code from visual layouts. Video training aligns events with timestamps for locating moments in a clip. Provider: Alibaba. Released 2025-10-15. Access: Open.
Aggregate Rating 1588 ± 4, rank #480 of 1656 rated models, from 621 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| DAR (Dynamic Affective Reasoning) - Affective Segmentation | 48.2 | Mean temporal IoU (%; overlap between predicted and ground-t | 100 |
| DAR (Dynamic Affective Reasoning) - Emotion Accuracy | 17.9 | Emotion accuracy (%; a segment counts only if its predicted | 100 |
| DAR (Dynamic Affective Reasoning) - Reasoning Quality | 2.7 | GPT-Score (out of 5): GPT-4o judge rating of each rationale | 100 |
| FGTR-Bench | 70.1 | Overall accuracy (%) on the 4,947-question blind test split, | 100 |
| FGTR-Bench - Holistic Sign | 86.8 | Accuracy (%) on 1,447 holistic traffic-sign questions, 4-opt | 100 |
| FGTR-Bench - Nighttime Signal | 83.2 | Accuracy (%) on 750 nighttime traffic-signal questions, 4-op | 100 |
| Flat-Pack Bench - Part Mating | 59.77 | Mate: decide whether two parts are connected in the final as | 100 |
| MinCU - Object | 69.19 | Grounded F1 (%; object changes (an entity replaced, introduc | 100 |
| NutriBench-Kitchen - Memory Management (Hard) | 70.7 | Score (%; the memory management task family, hard regime: di | 100 |
| OpenRef - Single Target | 75.2 | Grounding F1 (%) on expressions that refer to one object, ze | 100 |
| PInVerify (Attribute Decomposition) | 0.85 | Verification accuracy (0-1; share of 3,000 episodes, 1,000 e | 100 |
| RoadSafe365 - Road Type | 74.3 | Four-option multiple-choice accuracy (%) on the road-type qu | 100 |
Interactive version: theaggregate.ai/model?slug=qwen-3-vl-4b · How It Works · Data refreshed daily, snapshot 2026-10-20.