Aquila-VL-2B: benchmark results
BAAI's 2B vision-language model built on the LLaVA-OneVision framework, pairing a Qwen2.5-1.5B-Instruct backbone with a SigLIP vision tower, trained on 40 million image-text pairs. Provider: BAAI. Released 2024-10-25. Access: Open.
Aggregate Rating 1395 ± 12, rank #1238 of 1656 rated models, from 1070 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| OpenVLM MMT-Bench - Sketch-to-Code | 70 | Score (%) | 99.5 |
| OpenVLM MMT-Bench - Other Biological Attributes | 90 | Score (%) | 97.8 |
| OpenVLM MMT-Bench - Scene Recognition | 85 | Score (%) | 97.3 |
| MEGA-Bench Task - MMSoc Memotion | 68.2 | Task Score (%) | 96.5 |
| MEGA-Bench Task - Paper Review Acceptance | 60 | Task Score (%) | 96.5 |
| OpenVLM MMT-Bench - Tech Engineering | 55.6 | Score (%) | 96.4 |
| OpenVLM MMBench V1.1 EN - Image Topic | 98.9 | Accuracy (%) | 95.2 |
| OpenVLM MMT-Bench - Animals Recognition | 100 | Score (%) | 94.9 |
| OpenVLM Q-Bench - Yes-or-No (In-context Others) | 91.4 | Accuracy (%) | 94.3 |
| MEGA-Bench Task - MVSA Sentiment Classification | 78.6 | Task Score (%) | 94.2 |
| MEGA-Bench Task - Video Grounding Temporal | 40 | Task Score (%) | 94.2 |
| OpenVLM MMT-Bench - Fashion Recognition | 85 | Score (%) | 94.2 |
Interactive version: theaggregate.ai/model?slug=aquila-vl-2b · How It Works · Data refreshed daily, snapshot 2026-10-20.