QIMMA - ArabicMMLU: leaderboard
QIMMA's filtered version of ArabicMMLU, native Arabic multiple-choice exam questions from school and university curricula; not the same run as the OALL or HELM ArabicMMLU boards. Scored with LightEval log-likelihood over the answer options with character-length normalization rather than generated answers, so reasoning-tuned models can land near chance; accuracy 0-100, higher is better.
Metric: Accuracy (%, log-likelihood multiple choice). Source: huggingface.co. Saturation forecast: Around September 2027. 83 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Jais-2-70B-Chat | 81.29 | #430 |
| 2 | Qwen 3.5 397B A17B | 77.54 | #154 |
| 3 | Gemma 4 31B | 77.25 | #212 |
| 4 | Jais-2-8B-Chat | 73.96 | #802 |
| 5 | Qwen 2.5 72B Instruct | 73.78 | #469 |
| 6 | Qwen 3.5 122B A10B | 73.17 | #184 |
| 7 | ALLaM-7B-Instruct-preview | 72.54 | #724 |
| 8 | gemma-3-27B-pt | 72.27 | #505 |
| 9 | AIC-1 | 72.02 | #235 |
| 10 | Llama 3.3 70B Instruct | 71.57 | #541 |
| 11 | Qwen 3.5 35B A3B Base | 71.52 | #319 |
| 12 | Gemma 4 26B A4B | 71.14 | #360 |
| 13 | AceGPT-v2-32B-Chat | 70.62 | #518 |
| 14 | Qwen 3.5 27B | 69.47 | #192 |
| 15 | Qwen 2.5 32B Instruct | 68.76 | #509 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=qimma-arabicmmlu · How It Works · Data refreshed daily, snapshot 2026-11-05.