Qwen 3 VL 4B Instruct: 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 1579 ± 5, rank #515 of 1656 rated models, from 653 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| CVSBench - CVUSA Satellite-to-Ground | 62.4 | Accuracy (%) on CVUSA-subset satellite-to-ground questions: | 100 |
| EgoPro-Bench - Action | 52.24 | F1 (%; alert timing for scene events and human actions, 200 | 100 |
| Embodied3DBench - 2D Average | 56.5 | Mean of the five 2D task scores, each a normalized score (0- | 100 |
| Embodied3DBench - Correspondence (2D) | 58.3 | Correspondence (2D) task (cross-view correspondence of a giv | 100 |
| Embodied3DBench - Correspondence (3D) | 42.1 | Correspondence (3D) task (cross-view correspondence of a giv | 100 |
| Embodied3DBench - Grasp Point (2D) | 84.9 | Grasp Point (2D) task (grasp points in the image plane), nor | 100 |
| Embodied3DBench - Manipulation Trajectory (2D) | 37.6 | Manipulation Trajectory (2D) task (end-effector key trajecto | 100 |
| Embodied3DBench - Manipulation Trajectory (3D) | 44.5 | Manipulation Trajectory (3D) task (end-effector key trajecto | 100 |
| EviSafe - Natural Severity Accuracy | 52.8 | Accuracy (%) | 100 |
| VLM-RobustBench (MMMU-Pro) | -6.8 | Mean relative corruption error (percent of visual gain lost) | 100 |
| DatBench - Scene OCR | 77.9 | Accuracy (%) | 96.2 |
| DatBench - Document Understanding | 71 | Accuracy (%) | 96 |
Interactive version: theaggregate.ai/model?slug=qwen-3-vl-4b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.