TurkBench - NLI: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 256 multiple-choice natural language inference items (entailment, contradiction or neutral) built from Turkish sociology texts. 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 27B | 86.67 | #192 |
| 2 | Qwen 3.5 397B A17B | 85.49 | #154 |
| 3 | GLM-4.6 | 85.49 | #265 |
| 4 | Qwen 3 Next 80B A3B Instruct | 85.1 | #381 |
| 5 | Qwen 3.5 122B A10B | 83.53 | #184 |
| 6 | Qwen 3.5 35B A3B | 82.75 | #268 |
| 7 | Qwen 2.5 14B Instruct | 82.35 | #654 |
| 8 | Qwen 3 32B | 81.56 | #449 |
| 9 | Qwen 3 30B A3B 2507 Instruct | 81.56 | #492 |
| 10 | GLM-5 FP8 | 81.18 | #147 |
| 11 | GPT-OSS-120B | 80 | #346 |
| 12 | MiniMax-M2.5 | 76.86 | #316 |
| 13 | Gemma 2 9B (IT) | 76.47 | #795 |
| 14 | GLM-4.7 Flash | 76.47 | #543 |
| 15 | Gemma 3 27B (IT) | 75.69 | #528 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=turkbench-nli · How It Works · Data refreshed daily, snapshot 2026-11-05.