Qwen 3 VL 32B (Thinking): benchmark results
Alibaba Qwen 3 VL 32B vision-language model evaluated with thinking enabled. Provider: Alibaba. Released 2025-10-21. Access: Open.
Aggregate Rating 1599 ± 5, rank #716 of 2168 rated models, from 445 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AeroGround - Region Counting | 45.3 | Accuracy (%; Spatial Understanding: count category instances | 100 |
| DefectBench (Building Facades) | 84.42 | Defect identification F1 (0-1, times 100) of the primary def | 100 |
| ERQA-Plus | 83.4 | Multiple-choice accuracy (%) over all 1,766 questions in 12 | 100 |
| ERQA-Plus - Action-Centric Reasoning | 84.5 | Multiple-choice accuracy (%) on the 323 action-centric reaso | 100 |
| ERQA-Plus - Contextual and Commonsense Reasoning | 88.6 | Multiple-choice accuracy (%) on the 297 contextual and commo | 100 |
| ERQA-Plus - Navigation and Environmental Reasoning | 80.1 | Multiple-choice accuracy (%) on the 171 navigation and envir | 100 |
| ERQA-Plus - Open-Ended | 61.4 | Mean Sentence-BERT cosine similarity (x100, -100 to 100) bet | 100 |
| Flat-Pack Bench - Part Tracking | 45.53 | Track: recover the correspondence of shuffled part IDs betwe | 100 |
| LLM Stats (MuirBench) | 80.3 | Score (%) | 100 |
| LLM Stats (OCRBench-V2 (en)) | 68.4 | Score (%) | 100 |
| ReactBench (Hallucination) | 65.3 | React-Score (%): question-type-weighted accuracy (short answ | 100 |
| ReactBench (Hallucination) - Alteration Tracing (CoT) | 78.7 | Accuracy (%) on Alteration Tracing (10,956 questions on 913 | 100 |
Interactive version: theaggregate.ai/model?slug=qwen-3-vl-32b-thinking · How It Works · Data refreshed daily, snapshot 2026-10-20.