TurkBench - Math: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 500 mathematical reasoning problems from TÜBİTAK science olympiads and METU exams, each scored on the final answer. Accuracy 0-100, higher is better; some reasoning models score near zero.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around February 2027. 38 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5 FP8 | 81 | #147 |
| 2 | Qwen 3.5 35B A3B | 80.2 | #268 |
| 3 | Qwen 3.5 122B A10B | 80 | #184 |
| 4 | Qwen 3.5 397B A17B | 78.4 | #154 |
| 5 | DeepSeek V3.2 Exp | 65.8 | #241 |
| 6 | MiniMax-M2 | 63.6 | #327 |
| 7 | Qwen 3 Next 80B A3B Instruct | 63.4 | #381 |
| 8 | GLM-4.7 Flash | 59.8 | #543 |
| 9 | DeepSeek V3.1 | 58.6 | #275 |
| 10 | Qwen 3 30B A3B 2507 Instruct | 55.4 | #492 |
| 11 | Qwen 3.5 27B | 55.2 | #192 |
| 12 | MiniMax-M2.5 | 44.8 | #316 |
| 13 | MiMo-V2-Flash | 42.4 | #378 |
| 14 | Step 3.5 Flash | 38.4 | #281 |
| 15 | GLM-4.6 | 34.4 | #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-math · How It Works · Data refreshed daily, snapshot 2026-11-05.