Qwen 2.5 VL 32B Instruct: benchmark results
Alibaba Qwen 2.5 VL 32B vision-language instruction-tuned checkpoint. Provider: Alibaba. Released 2025-03-24. Access: Open.
Aggregate Rating 1597 ± 5, rank #453 of 1656 rated models, from 584 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| DefectBench (Building Facades) - Surface Stain Counting Error | 0.94 | Mean absolute error of the model's count of surface-stain in | 100 |
| EVALITA - MAIA-MC | 87.26 | CPS | 100 |
| EVALITA - word-in-context | 77.56 | CPS | 100 |
| LLM Stats (InfoVQA) | 83.4 | Score (%) | 100 |
| MMBU - Grounded Classification (Masks) | 69.3 | Closed-ended classification of a region marked by a segmenta | 100 |
| OrdinalBench (2D Single-Loop) | 32.83 | Final accuracy (%) of naming the N-th object when counting f | 100 |
| RebusBench (1-Shot) | 7.39 | Exact match: exact match accuracy (%) against the gold answe | 100 |
| RebusBench (3-Shot) | 8.08 | Exact match: exact match accuracy (%) against the gold answe | 100 |
| SAW-Bench - Self-Localization | 53 | Accuracy (%) on the 200 self-localization questions (where t | 100 |
| UVLM Street-View Benchmark (Reasoning) | 88 | Mean proximity score (%) in advanced-reasoning mode (chain o | 100 |
| UVLM Street-View Benchmark (Standard) | 83.3 | Mean proximity score (%) in standard mode (direct numeric an | 100 |
| Open Portuguese LLM - BLUEX | 80.53 | Accuracy (%) | 98.7 |
Interactive version: theaggregate.ai/model?slug=qwen-2-5-vl-32b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.