TurkBench - Named Entity Recognition: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 438 multiple-choice token-classification items identifying named entities in Turkish news text, run with thinking off. 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 397B A17B | 71.29 | #154 |
| 2 | Qwen 3.5 122B A10B | 70.24 | #184 |
| 3 | GPT-OSS-120B | 70.01 | #346 |
| 4 | GLM-4.6 | 69.19 | #265 |
| 5 | Step 3.5 Flash | 68.84 | #281 |
| 6 | Qwen 3.5 35B A3B | 68.3 | #268 |
| 7 | GLM-5 FP8 | 66.98 | #147 |
| 8 | Gemma 3 27B (IT) | 65.99 | #528 |
| 9 | MiniMax-M2.5 | 63.93 | #316 |
| 10 | Gemma 3 12B (IT) | 63.52 | #672 |
| 11 | MiniMax-M2 | 63.18 | #327 |
| 12 | DeepSeek V3.2 Exp | 62.05 | #241 |
| 13 | MiMo-V2-Flash | 61.1 | #378 |
| 14 | Qwen 3 32B | 60.4 | #449 |
| 15 | GLM-4.7 Flash | 60.28 | #543 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=turkbench-named-entity-recognition · How It Works · Data refreshed daily, snapshot 2026-11-05.