KLEAR - KMMLU-Redux - +K: leaderboard
KLEAR v1, KMMLU-Redux, 67 questions. Condition +K: C closed-book; +K adds anchor knowledge; +K+A adds application statements. One single-user-message attempt, step-by-step instruction, answer extracted after Korean Answer marker; unextractable answers count incorrect. Exact endpoint, inference access, reasoning, sampling and max output tokens are pinned from Table21; API developer endpoints pinned by authors. Three conditions are not pooled.
Metric: Score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | claude-sonnet-5-5 (KLEAR Run: endpoint=claude-sonnet-5-5;access=Anthropic;reasoning=High;sampling=default;max_output=32000) | 95.5 | |
| 2 | gpt-6.1-sol (KLEAR Run: endpoint=gpt-6.1-sol;access=OpenAI;reasoning=High;sampling=default;max_output=32000) | 95.5 | |
| 3 | glm-4.7 (KLEAR Run: endpoint=z-ai/glm-4.7;access=OpenRouter;reasoning=High;sampling=default;max_output=32000) | 91 | |
| 4 | qwen3.5-397b-a17b (KLEAR Run: endpoint=qwen/qwen3.5-397b-a17b;access=OpenRouter;reasoning=High;sampling=default;max_output=32000) | 89.6 | |
| 5 | K-EXAONE-236B-A23B (KLEAR Run: endpoint=LGAI-EXAONE/K-EXAONE-236B-A23B;access=vLLM;reasoning=Enabled;sampling=1.0 / 0.95;max_output=28672) | 88.1 | |
| 6 | A.X-K1 (KLEAR Run: endpoint=skt/A.X-K1;access=vLLM;reasoning=Enabled;sampling=0.6 / 0.95;max_output=45056) | 74.6 | |
| 7 | Solar-Open-100B (KLEAR Run: endpoint=upstage/Solar-Open-100B;access=vLLM;reasoning=High;sampling=0.8 / 0.95;max_output=28672) | 74.6 | |
| 8 | c4ai-aya-expanse-32b (KLEAR Run: endpoint=c4ai-aya-expanse-32b;access=Cohere;reasoning=None;sampling=0.3;max_output=4000) | 35.8 |
Interactive version: theaggregate.ai/benchmark?slug=klear-kmmlu-redux-plusk · How It Works · Data refreshed daily, snapshot 2026-11-12.