KRETA - System 1: leaderboard

The 1,426 KRETA items that only require reading and understanding the Korean text in the image; the score is the share answered correctly. It isolates text recognition from the reasoning items of System 2.

Metric: Accuracy on 1,426 System 1 (text recognition and understanding) 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)95.9#864
2A.X-4.0-VL-Light95.3#2575
3Qwen 2.5 VL 7B Instruct94.5#1893
4Qwen 2.5 VL 3B Instruct94.2#2648
5kanana-1.5-v-3B-instruct94#3953
6VARCO-VISION-2.0-14B93.5#1794
7Claude 3.5 Sonnet (20241022)93.4#648
8VARCO-VISION-14B90.9#1854
9InternVL2.5-4B90.7#2559
10InternVL2.5-8B89.8#2401
11MiniCPM-o-2.684.1#2417
12Ovis1.6-Gemma2-9B68.9#2365
13DeepSeek-VL2-Small67.3#2502
14Ovis1.6-Llama3.2-3B62.8#3750
15DeepSeek-VL2-Tiny60.8#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-system-1 · How It Works · Data refreshed daily, snapshot 2026-11-17.