OpenVLM TextVQA Val: leaderboard
TextVQA validation split from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): questions that require reading text in natural images, scored with the VQA accuracy metric in percent, higher is better. Exact-zero rows are failed runs and are dropped.
Metric: VQA Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 120 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2 VL 72B | 85.5 | #450 |
| 2 | InternVL3-14B | 82.2 | #506 |
| 3 | InternVL3-8B | 82.1 | #624 |
| 4 | MiniCPM-V-2.6 | 78.3 | #827 |
| 5 | Aquila-VL-2B | 76.5 | #1216 |
| 6 | Pixtral-12B | 75.7 | #805 |
| 7 | GPT-4.1 (2025-04-14) | 71 | #257 |
| 8 | Gemma 3 27B | 66.7 | #606 |
| 9 | Gemma 3 12B | 66.5 | #674 |
| 10 | GPT-4.1 Mini | 64.9 | #363 |
| 11 | Gemini 1.5 Flash | 61.6 | #620 |
| 12 | GPT-4.1 Nano | 60.9 | #730 |
| 13 | Gemma 3 4B | 57.7 | #1088 |
| 14 | Gemini 1.0 Pro | 36.9 | #900 |
| 15 | Qwen-VL-Max | 8.6 | #475 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-textvqa-val · How It Works · Data refreshed daily, snapshot 2026-11-05.