TurkBench - Commonsense Reasoning: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 241 multiple-choice commonsense inference items grounded in Turkish cultural knowledge and everyday situations. Accuracy 0-100, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 35B A3B | 95.08 | #268 |
| 2 | Qwen 3.5 122B A10B | 95.08 | #184 |
| 3 | GPT-OSS-120B | 94.67 | #346 |
| 4 | Qwen 3.5 27B | 93.44 | #192 |
| 5 | GLM-4.7 Flash | 93.44 | #543 |
| 6 | Step 3.5 Flash | 93.44 | #281 |
| 7 | Qwen 3.5 397B A17B | 93.03 | #154 |
| 8 | MiniMax-M2.5 | 93.03 | #316 |
| 9 | Qwen 3 Next 80B A3B Instruct | 92.21 | #381 |
| 10 | DeepSeek V3.1 | 91.8 | #275 |
| 11 | Gemma 3 27B (IT) | 91.39 | #528 |
| 12 | Gemma 3 12B (IT) | 91.39 | #672 |
| 13 | Qwen 3 32B | 90.98 | #449 |
| 14 | Qwen 3 30B A3B 2507 Instruct | 90.57 | #492 |
| 15 | Qwen 2.5 14B Instruct | 90.16 | #654 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=turkbench-commonsense-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-05.