TurkBench - STS: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 225 Turkish sentence pairs scored for semantic similarity, run with thinking off; the score is the Pearson correlation with gold ratings multiplied by 100, so it can be negative. Higher is better.
Metric: Pearson correlation x100. Source: huggingface.co. Saturation forecast: Estimated already saturated. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5 FP8 | 93.55 | #147 |
| 2 | Qwen 3.5 397B A17B | 93.34 | #154 |
| 3 | Qwen 3 Next 80B A3B Instruct | 92.95 | #381 |
| 4 | DeepSeek V3.2 Exp | 92.71 | #241 |
| 5 | Qwen 3.5 27B | 92.6 | #192 |
| 6 | Gemma 3 12B (IT) | 92.59 | #672 |
| 7 | Qwen 2.5 14B Instruct | 91.86 | #654 |
| 8 | Qwen 3.5 35B A3B | 91.83 | #268 |
| 9 | Qwen 2.5 7B Instruct | 91.8 | #863 |
| 10 | DeepSeek V3.1 | 91.54 | #275 |
| 11 | Qwen 3.5 122B A10B | 91.31 | #184 |
| 12 | Qwen 3 32B | 90.94 | #449 |
| 13 | MiMo-V2-Flash | 90.42 | #378 |
| 14 | GPT-OSS-120B | 90.41 | #346 |
| 15 | Gemma 3 27B (IT) | 90.17 | #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-sts · How It Works · Data refreshed daily, snapshot 2026-11-05.