OpenVLM MTVQA: leaderboard
OpenCompass OpenVLM evaluation of MTVQA: multilingual text-centric visual question answering over images containing text in Arabic, German, French, Italian, Japanese, Korean, Russian, Thai, Vietnamese, and Chinese.
Metric: Accuracy (%). Source: huggingface.co. Status: saturated. 157 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-4.1 Mini | 36.8 |
| 2 | GPT-4.1 (2025-04-14) | 34.1 |
| 3 | InternVL3-78B | 32.5 |
| 4 | InternVL3-38B | 32.4 |
| 5 | Gemini 1.5 Pro | 32 |
| 6 | Claude 3.7 Sonnet | 30.6 |
| 7 | Gemini 1.5 Flash | 30.5 |
| 8 | InternVL3-8B | 30.4 |
| 9 | Qwen 2 VL 72B | 29.6 |
| 10 | Gemma 3 27B | 28.9 |
| 11 | GPT-4.1 Nano | 27.6 |
| 12 | Qwen 2 VL 7B | 25.8 |
| 13 | Gemma 3 12B | 23.8 |
| 14 | Grok 2 (1212) | 22.6 |
| 15 | Gemma 3 4B | 22.1 |
Interactive version: theaggregate.ai/benchmark?slug=openvlm-mtvqa · How It Works · Data refreshed daily, snapshot 2026-09-05.