OpenVLM ScienceQA Val: leaderboard
ScienceQA validation split from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): multiple-choice science questions with image context from school curricula. Accuracy in percent, higher is better; the test split is the separate OpenVLM ScienceQA Test board.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 218 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | InternVL3-38B | 99.2 | #413 |
| 2 | InternVL3-78B | 98.5 | #361 |
| 3 | InternVL3-14B | 98.4 | #506 |
| 4 | InternVL3-8B | 97.9 | #624 |
| 5 | InternVL2-8B | 97.4 | #813 |
| 6 | MiniCPM-V-2.6 | 96 | #827 |
| 7 | Aquila-VL-2B | 93.7 | #1216 |
| 8 | GPT-4.1 (2025-04-14) | 92.6 | #257 |
| 9 | Qwen 2 VL 72B | 91.6 | #450 |
| 10 | Claude 3.7 Sonnet | 90.9 | #259 |
| 11 | Grok 2 (1212) | 90.3 | #502 |
| 12 | Claude 3.5 Sonnet | 89.3 | #362 |
| 13 | Qwen 2.5 VL 7B | 88.9 | #641 |
| 14 | Gemma 3 27B | 87.1 | #606 |
| 15 | Pixtral-12B | 86.6 | #805 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-scienceqa-val · How It Works · Data refreshed daily, snapshot 2026-11-05.