TurkBench - MMLU Topics: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 2,373 Turkish multiple-choice questions on humanities, social sciences and STEM drawn from ÖSYM exams (YGS, LYS, KPSS, TUS) and METU course exams. Accuracy 0-100, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around February 2027. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 80.38 | #154 |
| 2 | Qwen 3.5 122B A10B | 77.75 | #184 |
| 3 | Qwen 3.5 35B A3B | 77.2 | #268 |
| 4 | GLM-4.6 | 76.02 | #265 |
| 5 | MiniMax-M2.5 | 75.84 | #316 |
| 6 | Qwen 3.5 27B | 75.66 | #192 |
| 7 | Step 3.5 Flash | 72.66 | #281 |
| 8 | GPT-OSS-120B | 72.2 | #346 |
| 9 | MiniMax-M2 | 68.66 | #327 |
| 10 | GLM-4.7 Flash | 67.39 | #543 |
| 11 | GLM-5 FP8 | 61.85 | #147 |
| 12 | Qwen 3 Next 80B A3B Instruct | 61.4 | #381 |
| 13 | Qwen 3 30B A3B 2507 Instruct | 58.94 | #492 |
| 14 | Gemma 3 27B (IT) | 53.5 | #528 |
| 15 | Gemma 3 12B (IT) | 48.23 | #672 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=turkbench-mmlu-topics · How It Works · Data refreshed daily, snapshot 2026-11-05.