Qwen 3 VL 2B (Thinking): benchmark results
Provider: Alibaba. Released 2025-10-21. Access: Open.
Aggregate Rating 1490 ± 14, rank #2514 of 10916 rated models, from 160 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| DatBench - Math & Logic | 25.7 | Accuracy (%) | 73.1 |
| DatBench - Chart Understanding | 62.3 | Accuracy (%) | 69.2 |
| DatBench - General VQA | 51.3 | Accuracy (%) | 69.2 |
| DatBench - Counting | 91.2 | Accuracy (%) | 65.4 |
| EXPLORE-Bench - Object Coverage | 57.26 | Object-level coverage (0-100): soft Sentence-BERT matching o | 63 |
| MedLayXPlain | 59.1 | S (self-reported) | 59.7 |
| HAERAE-Vision - Mathematics | 26.97 | Checklist Score (%; original under-specified queries, GPT-5- | 59.1 |
| FailBench - reassemble | 51 | Balanced Accuracy (%) | 58.3 |
| EXPLORE-Bench - Relation Score | 2.56 | Relation-level score (0-5): a Qwen3-8B scorer rates the pred | 55.6 |
| EPIC-Bench - Affordance Region | 6.07 | Manipulation score (0-100) on affordance region localization | 52.8 |
| HAERAE-Vision - Entertainment/Arts | 9.03 | Checklist Score (%; original under-specified queries, GPT-5- | 50 |
| EXPLORE-Bench - Short Sequences | 46.13 | Unified score (0-100): 0.25 x object coverage + 0.35 x attri | 48.1 |
Interactive version: theaggregate.ai/model?slug=qwen-3-vl-2b-thinking · How It Works · Data refreshed daily, snapshot 2026-11-08.