Qwen 3 VL 4B (Thinking): benchmark results
Alibaba Qwen 3 VL 4B vision-language model evaluated with thinking enabled. Provider: Alibaba. Released 2025-10-15. Access: Open.
Aggregate Rating 1508 ± 7, rank #1045 of 2168 rated models, from 311 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Flat-Pack Bench - Part Mating | 59.77 | Mate: decide whether two parts are connected in the final as | 100 |
| DatBench - Math & Logic | 38.2 | Accuracy (%) | 92.3 |
| FailBench - botfails | 68 | Balanced Accuracy (%) | 91.7 |
| MultihopSpatial - 2-Hop Exo-Centric | 63.9 | Multiple-choice accuracy (%) on the 2-hop exo-centric questi | 88.9 |
| DatBench - Chart Understanding | 68.9 | Accuracy (%) | 88.5 |
| DatBench - General VQA | 59.8 | Accuracy (%) | 88.5 |
| FailBench - roboarena | 72 | Balanced Accuracy (%) | 87.5 |
| CSR-Bench (Hallucination) - Distractor Text | 40.7 | Alignment accuracy (%) when fabricated common-knowledge text | 86.7 |
| ChartArena - Box Plot (Chinese) | 29 | mAP-high (%) on Chinese box plots, each model parses the cha | 83.3 |
| OCR-Robust | 78.63 | OCR1.0 Clean (self-reported) | 82.4 |
| MultihopSpatial - 3-Hop Ego-Centric | 20.4 | Multiple-choice accuracy (%) on the 3-hop ego-centric questi | 81.9 |
| ChartArena - Box Plot (English) | 22.1 | mAP-high (%) on English box plots, each model parses the cha | 81.4 |
Interactive version: theaggregate.ai/model?slug=qwen-3-vl-4b-thinking · How It Works · Data refreshed daily, snapshot 2026-10-20.