QIMMA - MizanQA: leaderboard
MizanQA multiple-choice questions on Moroccan law, as run in the QIMMA Arabic leaderboard. The score is LightEval's normalized multiple-choice probability metric, computed from option log-likelihoods rather than generated answers; 0-100, higher is better.
Metric: Normalized multiple-choice probability score (%). Source: huggingface.co. Saturation forecast: Around 2030. 83 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 73.36 | #154 |
| 2 | Jais-2-70B-Chat | 71.78 | #430 |
| 3 | Llama 3.3 70B Instruct | 67.44 | #541 |
| 4 | Qwen 3.5 27B | 65.18 | #192 |
| 5 | Gemma 4 31B (IT) | 64.02 | #178 |
| 6 | Qwen 3.5 122B A10B | 64.01 | #184 |
| 7 | AceGPT-v2-32B-Chat | 63.89 | #518 |
| 8 | Qwen 2.5 72B Instruct | 63.49 | #469 |
| 9 | Gemma 3 27B (IT) | 61.65 | #528 |
| 10 | Jais-2-8B-Chat | 61.33 | #802 |
| 11 | c4ai-command-r7B-arabic-02-2025 | 59.98 | #727 |
| 12 | Atlas-Chat-9B | 59.69 | #710 |
| 13 | Qwen 3.6 35B A3B | 59.4 | #238 |
| 14 | Qwen 2.5 7B Instruct | 58.91 | #863 |
| 15 | Qwen 3.5 35B A3B | 58.68 | #268 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=qimma-mizanqa · How It Works · Data refreshed daily, snapshot 2026-11-05.