QIMMA - 3LM STEM: leaderboard
Native Arabic STEM multiple-choice questions from TII's 3LM benchmark (science and mathematics), as run in the QIMMA Arabic leaderboard. 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: Not forecast. 83 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 35B A3B Base | 95.33 | #319 |
| 2 | Gemma 4 31B | 91.73 | #212 |
| 3 | gemma-3-27B-pt | 90 | #505 |
| 4 | Qwen 3.5 397B A17B | 88.67 | #154 |
| 5 | Llama 3.3 70B Instruct | 88.28 | #541 |
| 6 | AIC-1 | 88.13 | #235 |
| 7 | Jais-2-70B-Chat | 87.96 | #430 |
| 8 | Qwen 2.5 72B Instruct | 87.55 | #469 |
| 9 | Qwen3.5-4B-Base | 87.25 | #799 |
| 10 | Qwen 3.5 122B A10B | 86.18 | #184 |
| 11 | Qwen3-8B-Base | 85.03 | #626 |
| 12 | Gemma 4 26B A4B | 84.95 | #360 |
| 13 | AceGPT-v2-32B-Chat | 84.88 | #518 |
| 14 | Gemma 3 27B (IT) | 84.75 | #528 |
| 15 | Hala-9B | 84.49 | #517 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=qimma-3lm-stem · How It Works · Data refreshed daily, snapshot 2026-11-05.