OpenVLM MTVQA - Japanese: leaderboard
MTVQA Japanese accuracy from the OpenCompass OpenVLM Leaderboard (VLMEvalKit): questions about Japanese text in scene and document images, asked and answered in Japanese. One of the nine language splits of the multilingual text-centric VQA benchmark; accuracy in percent, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 157 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 Mini | 24 | #363 |
| 2 | InternVL3-8B | 22.3 | #624 |
| 3 | InternVL3-78B | 22.3 | #361 |
| 4 | GPT-4.1 (2025-04-14) | 22.2 | #257 |
| 5 | InternVL3-38B | 21.5 | #413 |
| 6 | InternVL3-14B | 21.5 | #506 |
| 7 | Claude 3.7 Sonnet | 21.5 | #259 |
| 8 | Qwen 2 VL 72B | 19.8 | #450 |
| 9 | Gemini 1.5 Pro | 18.8 | #454 |
| 10 | Qwen 2.5 VL 7B | 18.5 | #641 |
| 11 | Gemma 3 27B | 17.6 | #606 |
| 12 | Gemini 1.5 Flash | 17.1 | #620 |
| 13 | GPT-4.1 Nano | 17 | #730 |
| 14 | Qwen 2 VL 7B | 16.8 | #743 |
| 15 | Moonshot v1 8K | 15.1 | #622 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-mtvqa-japanese · How It Works · Data refreshed daily, snapshot 2026-11-05.