KRETA - System 2: leaderboard
The 1,151 KRETA items that need domain knowledge or multi-step reasoning over the Korean text in the image, including adapted CSAT science and history questions; the score is the share answered correctly.
Metric: Accuracy on 1,151 System 2 (knowledge and multi-step reasoning) items (%). Source: github.com. Saturation forecast: Estimated already saturated. 17 models tracked. Updated 17 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-11-20) | 70.5 | #864 |
| 2 | Claude 3.5 Sonnet (20241022) | 64.5 | #648 |
| 3 | A.X-4.0-VL-Light | 56.5 | #2575 |
| 4 | VARCO-VISION-2.0-14B | 53.1 | #1794 |
| 5 | kanana-1.5-v-3B-instruct | 51.4 | #3953 |
| 6 | VARCO-VISION-14B | 49.3 | #1854 |
| 7 | InternVL2.5-8B | 47.3 | #2401 |
| 8 | InternVL2.5-4B | 45.9 | #2559 |
| 9 | Ovis1.6-Gemma2-9B | 45.4 | #2365 |
| 10 | Qwen 2.5 VL 3B Instruct | 43.9 | #2648 |
| 11 | MiniCPM-o-2.6 | 39.9 | #2417 |
| 12 | Ovis1.6-Llama3.2-3B | 39.1 | #3750 |
| 13 | Qwen 2.5 VL 7B Instruct | 36.1 | #1893 |
| 14 | DeepSeek-VL2-Small | 36.1 | #2502 |
| 15 | DeepSeek-VL2-Tiny | 34 | #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-2 · How It Works · Data refreshed daily, snapshot 2026-11-17.