TurkBench - Truthfulness: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 250 expert-written Turkish questions targeting common misconceptions and folk wisdom, graded by an LLM judge (0-100, higher is better).
Metric: LLM-as-a-judge score (0-100). Source: huggingface.co. Saturation forecast: Estimated already saturated. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 122B A10B | 91.79 | #184 |
| 2 | Qwen 3.5 27B | 91.39 | #192 |
| 3 | Qwen 3.5 397B A17B | 91.08 | #154 |
| 4 | GLM-5 FP8 | 90.54 | #147 |
| 5 | Qwen 3.5 35B A3B | 89.24 | #268 |
| 6 | MiniMax-M2 | 88.03 | #327 |
| 7 | MiniMax-M2.5 | 87.85 | #316 |
| 8 | DeepSeek V3.2 Exp | 83.72 | #241 |
| 9 | Step 3.5 Flash | 82.11 | #281 |
| 10 | MiMo-V2-Flash | 81.61 | #378 |
| 11 | DeepSeek V3.1 | 77.21 | #275 |
| 12 | GLM-4.7 Flash | 77.13 | #543 |
| 13 | GPT-OSS-120B | 75.33 | #346 |
| 14 | Qwen 3 Next 80B A3B Instruct | 67.89 | #381 |
| 15 | GLM-4.6 | 67.04 | #265 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=turkbench-truthfulness · How It Works · Data refreshed daily, snapshot 2026-11-05.