TurkBench - Faithfulness: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 250 expert-written items requiring answers grounded in a supplied Turkish context, as in retrieval-augmented generation, 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 | 90.12 | #184 |
| 2 | Qwen 3.5 27B | 89.22 | #192 |
| 3 | Qwen 3 Next 80B A3B Instruct | 88.71 | #381 |
| 4 | MiniMax-M2.5 | 88.58 | #316 |
| 5 | Qwen 3.5 397B A17B | 88.23 | #154 |
| 6 | Gemma 3 12B (IT) | 88.14 | #672 |
| 7 | DeepSeek V3.2 Exp | 88.14 | #241 |
| 8 | MiniMax-M2 | 88.02 | #327 |
| 9 | Step 3.5 Flash | 87.99 | #281 |
| 10 | Qwen 3.5 35B A3B | 87.97 | #268 |
| 11 | GLM-5 FP8 | 87.93 | #147 |
| 12 | GLM-4.6 | 87.84 | #265 |
| 13 | Qwen 3 32B | 87.56 | #449 |
| 14 | DeepSeek V3.1 | 87.36 | #275 |
| 15 | aya-expanse-8B | 86.98 | #914 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=turkbench-faithfulness · How It Works · Data refreshed daily, snapshot 2026-11-05.