KRETA: leaderboard

Vision-language models answer 2,577 Korean multiple-choice questions about text-rich images (signs, documents, menus, exam pages) across 15 domains; every model ran all items and the score is the share answered correctly. Unlike English text-VQA sets, the text, domains and the CSAT exam items are Korean.

Metric: Accuracy on all 2,577 Korean text-rich VQA items (%). Source: github.com. Saturation forecast: Estimated already saturated. 17 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1GPT-4o (2024-11-20)84.6#864
2Claude 3.5 Sonnet (20241022)80.5#648
3A.X-4.0-VL-Light78#2575
4VARCO-VISION-2.0-14B75.4#1794
5kanana-1.5-v-3B-instruct75#3953
6VARCO-VISION-14B72.3#1854
7Qwen 2.5 VL 3B Instruct71.8#2648
8InternVL2.5-8B70.8#2401
9InternVL2.5-4B70.7#2559
10Qwen 2.5 VL 7B Instruct68.5#1893
11MiniCPM-o-2.664.3#2417
12Ovis1.6-Gemma2-9B58.4#2365
13DeepSeek-VL2-Small53.3#2502
14Ovis1.6-Llama3.2-3B52.2#3750
15DeepSeek-VL2-Tiny48.8#6676

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=kreta · How It Works · Data refreshed daily, snapshot 2026-11-17.