Qwen 2.5 VL 3B: benchmark results
Provider: Alibaba. Released 2025-01-26. Access: Open.
Aggregate Rating 1482 ± 10, rank #2635 of 10916 rated models, from 702 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| PaveVQA - Classification Accuracy | 65.18 | Classification accuracy (%) on categorical questions about d | 100 |
| PaveVQA - Quantification MAPE | 116.2 | Mean absolute percentage error (%, no upper limit) of quanti | 100 |
| OpenVLM Q-Bench - What (In-context Others) | 92.2 | Accuracy (%) | 98.4 |
| OpenVLM MMT-Bench - Image Quality Assessment | 70 | Score (%) | 96.8 |
| OpenVLM MMBench V1.1 CN - Image Scene | 92.4 | Accuracy (%) | 96.6 |
| OpenVLM MMT-Bench - Pixel Localization | 50 | Score (%) | 96.4 |
| OpenVLM MMBench V1.1 CN - Celebrity Recognition | 92.5 | Accuracy (%) | 95.2 |
| OpenVLM MMBench V1.1 CN - Social Relation | 90.1 | Accuracy (%) | 95.2 |
| OpenVLM MMT-Bench - Equation-to-LaTeX | 95 | Score (%) | 95.1 |
| PolicyShiftBench | 61 | F1 of the block decision (%; policy-conditioned image guardr | 94.4 |
| PolicyShiftBench - Adaptive Split | 65.9 | F1 of the block decision (%; policy-conditioned image guardr | 94.4 |
| OpenVLM Q-Bench - Yes-or-No (In-context Others) | 91.4 | Accuracy (%) | 94.3 |
Interactive version: theaggregate.ai/model?slug=qwen-2-5-vl-3b · How It Works · Data refreshed daily, snapshot 2026-11-08.