QIMMA - Overall: leaderboard
QIMMA (Technology Innovation Institute) Arabic LLM leaderboard average over 14 native-Arabic benchmarks spanning STEM, law, medicine, culture, poetry, safety and code, with MedArabiQ counted as the mean of its MCQ and QA parts. Open-weight models only; 0-100, higher is better. Multiple-choice sets are scored by log-likelihood, not generation.
Metric: Mean score across 14 benchmarks (%). Source: huggingface.co. Saturation forecast: Around September 2028. 83 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 69.24 | #154 |
| 2 | Qwen 2.5 72B Instruct | 68.49 | #469 |
| 3 | AIC-1 | 67.85 | #235 |
| 4 | Jais-2-70B-Chat | 67.77 | #430 |
| 5 | Llama 3.3 70B Instruct | 67.09 | #541 |
| 6 | Qwen 3.5 122B A10B | 66.33 | #184 |
| 7 | Qwen 2.5 32B Instruct | 66.09 | #509 |
| 8 | Gemma 3 27B (IT) | 63.84 | #528 |
| 9 | AceGPT-v2-32B-Chat | 62.96 | #518 |
| 10 | Gemma 3 12B (IT) | 62.65 | #672 |
| 11 | Qwen 3.5 27B | 62.01 | #192 |
| 12 | Qwen 3.5 35B A3B | 61.16 | #268 |
| 13 | Gemma 4 31B (IT) | 60.38 | #178 |
| 14 | Jais-2-8B-Chat | 59.33 | #802 |
| 15 | Qwen 2.5 14B Instruct | 59.29 | #654 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=qimma-overall · How It Works · Data refreshed daily, snapshot 2026-11-05.