DatBench - General VQA: leaderboard
DatBench (DatologyAI, 2026) general-VQA suite: holistic visual question answering and exam-style multimodal reasoning, drawn from MMMU-Pro, MMBench and VQA-v2 after DatBench's curation (multiple-choice items turned into free-form answers judged by Qwen3-30B or scored circularly, items solvable without the image or flagged by a GPT-5.2 judge as mislabeled removed, and a high-discrimination subset kept). Accuracy in percent, higher is better, with generation capped at 4,096 tokens; the curation used the responses of the same 1B-10B models the paper scores.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 27 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 8B Instruct | 63.7 | #982 |
| 2 | Qwen 3 VL 8B (Thinking) | 62.4 | #887 |
| 3 | Qwen 3 VL 4B Instruct | 59.9 | #1281 |
| 4 | Qwen 3 VL 4B (Thinking) | 59.8 | #1303 |
| 5 | Qwen 2.5 VL 7B Instruct | 56.8 | #1842 |
| 6 | GLM-4.1V-9B (Thinking) | 54.1 | #1061 (GLM-4.1V-9B) |
| 7 | Qwen2.5-Omni-7B | 41.7 | #1797 |
| 8 | InternVL2-8B | 34.9 | #3025 |
| 9 | InternVL2.5-2B | 27.9 | #6868 |
| 10 | Gemma 3 4B (IT) | 17.1 | #3848 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=datbench-general-vqa · How It Works · Data refreshed daily, snapshot 2026-11-12.