OpenVLM SEEDBench IMG - Visual Reasoning: leaderboard
SEED-Bench image Visual Reasoning accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): four-option multiple-choice questions on reasoning from visual evidence, such as commonsense inferences, one of the nine evaluation dimensions of the SEED-Bench image set. Accuracy in percent, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 226 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 VL 7B | 84 | #641 |
| 2 | GPT-4.1 (2025-04-14) | 83.1 | #257 |
| 3 | InternVL3-38B | 82.2 | #413 |
| 4 | Qwen-VL-Max | 82.2 | #475 |
| 5 | Qwen 2 VL 72B | 81.9 | #450 |
| 6 | InternVL3-8B | 81.6 | #624 |
| 7 | InternVL3-78B | 81.6 | #361 |
| 8 | Llama 3.2 90B Vision Instruct | 81.3 | #621 |
| 9 | GPT-4.1 Mini | 80.4 | #363 |
| 10 | InternVL3-14B | 80.4 | #506 |
| 11 | Grok 2 (1212) | 80.4 | #502 |
| 12 | Gemini 1.0 Pro | 80.4 | #900 |
| 13 | Qwen 2 VL 7B | 79.8 | #743 |
| 14 | Gemini 1.5 Pro | 79.5 | #454 |
| 15 | Aquila-VL-2B | 79.5 | #1216 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-seedbench-img-visual-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-05.