TurkBench - Sentiment Analysis: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 123 multiple-choice items classifying the sentiment of Turkish posts and reviews from sources such as X, Reddit, Amazon and Trendyol. Accuracy 0-100, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around March 2027. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 53.28 | #154 |
| 2 | GLM-5 FP8 | 51.09 | #147 |
| 3 | Qwen 3.5 27B | 50.36 | #192 |
| 4 | GLM-4.6 | 49.64 | #265 |
| 5 | Qwen 3.5 122B A10B | 48.91 | #184 |
| 6 | GPT-OSS-120B | 48.9 | #346 |
| 7 | Qwen 3 32B | 48.9 | #449 |
| 8 | DeepSeek V3.1 | 48.9 | #275 |
| 9 | MiniMax-M2 | 48.18 | #327 |
| 10 | Gemma 3 12B (IT) | 47.45 | #672 |
| 11 | Qwen 3.5 35B A3B | 47.45 | #268 |
| 12 | MiniMax-M2.5 | 47.45 | #316 |
| 13 | Step 3.5 Flash | 47.45 | #281 |
| 14 | Qwen 3 Next 80B A3B Instruct | 46.72 | #381 |
| 15 | Qwen 2.5 14B Instruct | 45.99 | #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-sentiment-analysis · How It Works · Data refreshed daily, snapshot 2026-11-05.