OpenVLM RealWorldQA: leaderboard
RealWorldQA accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): questions on real-world images, many taken from vehicles, testing spatial and physical understanding. Accuracy in percent, higher is better; exact-zero rows (failed runs) are dropped.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 235 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 (2025-04-14) | 78.7 | #257 |
| 2 | InternVL3-78B | 78.4 | #361 |
| 3 | GPT-4.1 Mini | 77.5 | #363 |
| 4 | Qwen 2 VL 72B | 76.7 | #450 |
| 5 | InternVL3-38B | 75.8 | #413 |
| 6 | Gemini 2.0 Flash | 72.3 | #351 |
| 7 | InternVL3-8B | 71.4 | #624 |
| 8 | InternVL3-14B | 69.8 | #506 |
| 9 | Gemini 1.5 Flash | 69 | #620 |
| 10 | Qwen 2 VL 7B | 68.5 | #743 |
| 11 | Qwen 2.5 VL 7B | 68.4 | #641 |
| 12 | Llama 3.2 90B Vision Instruct | 68.2 | #621 |
| 13 | Claude 3.5 Sonnet (20241022) | 65.8 | #302 |
| 14 | Pixtral-12B | 65.4 | #805 |
| 15 | MiniCPM-V-2.6 | 65 | #827 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-realworldqa · How It Works · Data refreshed daily, snapshot 2026-11-05.