TOPIK Decision - Reading Accuracy: leaderboard
Korean TOPIK II text-only reading questions; five decision models. The evaluation uses 502 text-only, four-choice TOPIK II reading questions from 12 officially released rounds. It excludes 72 image questions and 26 questions with missing passages from the original 600. Every model sees the same Korean question, passage and four choices in their original order, in independent requests without explanations. The denominator includes all 502 questions, including refusals. Hosted and local configurations retain their published model identities. Question access: Official exams are public. The source identifies the 12 rounds and exclusion rules but does not release the exact compiled 502-item subset.
Metric: Accuracy (%). Source: rickcarlino.com. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | JEV 1.14.0 (Benchmark Sweep Run: source=TOPIKDecision; hosted JEV; one request per item) | 97 | |
| 2 | GPT-6 Luna (Benchmark Sweep Run: source=TOPIKDecision; OpenAI Decisions hosted gpt-6-luna; two refusals count wrong) | 95 | |
| 3 | Clef Flash 9B (Benchmark Sweep Run: source=TOPIKDecision; Q8_0; Ollama 0.40.13; GPU; context 2K) | 92.4 | |
| 4 | Nimble 9B (Benchmark Sweep Run: source=TOPIKDecision; Q8_0; local; context about 8K) | 91.2 | |
| 5 | Tev1 4B (Benchmark Sweep Run: source=TOPIKDecision; Q8_0; local; context about 8K) | 85.3 |
Interactive version: theaggregate.ai/benchmark?slug=topik-decision-reading-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-06.