TurkBench - Instruction Following: leaderboard
TurkBench (METU NLP group and Turkcell, SIGTURK 2026) task: 997 expert-written Turkish instructions, from simple requests to multi-step tasks such as travel-cost estimation; responses are 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 | DeepSeek V3.2 Exp | 96.8 | #241 |
| 2 | DeepSeek V3.1 | 94.92 | #275 |
| 3 | Qwen 3 Next 80B A3B Instruct | 94.27 | #381 |
| 4 | GLM-5 FP8 | 94.08 | #147 |
| 5 | GPT-OSS-120B | 93.6 | #346 |
| 6 | Qwen 3.5 397B A17B | 93.42 | #154 |
| 7 | GLM-4.6 | 92.58 | #265 |
| 8 | Qwen 3 30B A3B 2507 Instruct | 92.51 | #492 |
| 9 | Qwen 3.5 35B A3B | 92.39 | #268 |
| 10 | Qwen 3.5 27B | 91.82 | #192 |
| 11 | Qwen 3.5 122B A10B | 91.54 | #184 |
| 12 | Step 3.5 Flash | 91.07 | #281 |
| 13 | MiMo-V2-Flash | 90.98 | #378 |
| 14 | MiniMax-M2 | 90.88 | #327 |
| 15 | MiniMax-M2.5 | 89.29 | #316 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=turkbench-instruction-following · How It Works · Data refreshed daily, snapshot 2026-11-05.