Qwen 3 VL 235B A22B Instruct: benchmark results
Can locate interface elements for GUI agents, read text in images and generate code from visual layouts. Video training aligns events with timestamps for locating moments in a clip. Provider: Alibaba. Released 2025-09-23. Access: Open.
Aggregate Rating 1669 ± 4, rank #273 of 1656 rated models, from 839 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AdaptMMBench (Oracle Visual Evidence) | 90.08 | Overall accuracy (%) on the 1,300 AdaptMMBench samples (auxi | 100 |
| AdaptMMBench - Key Step Coverage | 84.83 | Key step coverage (%): how far the adaptive-mode reasoning c | 100 |
| CFMME - Formula Recognition | 95.01 | Formula recognition CDM (%, character detection matching of | 100 |
| CFMME - Information Extraction | 93.43 | Field-level F1 (%) of key information extracted from financi | 100 |
| CFMME - Recognition, Detection and Extraction | 77.18 | Average (%) of seal recognition 1-NED, table recognition TED | 100 |
| CFMME - Seal Recognition | 88.27 | Seal text recognition score (%, 1 minus normalized edit dist | 100 |
| CFMME - Table Recognition | 78.92 | Table recognition TEDS (%, tree-edit-distance similarity of | 100 |
| ChronoVision - Artifact Dating | 49.94 | Accuracy (%) on the 887 Chinese historical artifact images ( | 100 |
| ENC-Bench - Point Features | 51.79 | Accuracy (%) on ENC-Bench's 7,830 point features questions ( | 100 |
| EmpathyBench | 51.3 | Average Score (%) | 100 |
| IDEAL-Bench - Room Type Accuracy | 90.3 | Room type accuracy (%) among parse-successful scenes: the pr | 100 |
| Konkur 1404 - Humanities | 52.15 | Accuracy (%, text-only) | 100 |
Interactive version: theaggregate.ai/model?slug=qwen-3-vl-235b-a22b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-20.