Jais 3 Model Cards - Dialectal Arabic MMLU (IFM-audited): leaderboard
MMLU-style questions in dialectal Arabic, on IFM's audited set. MBZUAI's Institute of Foundation Models reported this on the Jais 3 model cards (released October 15, 2026), comparing each Jais 3 size with size-matched reference models; this board collects all six cards. The cards state that all reported results use high reasoning effort. Vendor-reported. Metric: Accuracy (%). Higher is better. Source: https://huggingface.co/collections/IFM/jais-3-6abba9df4f1173ec7ce2929d
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3.8-Flash-Next | 90.1 | #69 |
| 2 | Qwen 3.5 397B A17B | 90 | #144 |
| 3 | Gemma 4 31B (IT) | 89.4 | #174 |
| 4 | Nemotron 3 Ultra | 88.7 | #234 |
| 5 | Qwen 3.8 27B | 87 | #104 |
| 6 | Muse Glimmer 30B | 87 | #178 |
| 7 | MiniMax-M3 | 86.6 | #137 |
| 8 | Gemma 4 12B (IT) | 83.5 | #371 |
| 9 | Qwen 3.5 9B | 80.4 | #377 |
| 10 | gemma-4-E4B-it | 76.6 | #566 |
| 11 | Qwen 3.5 4B | 76.5 | #473 |
| 12 | Falcon-H1-34B-Instruct | 71.5 | #666 |
| 13 | Gemma 4 E2B | 65.1 | #1076 |
| 14 | Qwen 3.5 2B | 59.5 | #1145 |
| 15 | Falcon-H1-7B-Instruct | 58.5 | #1043 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=jais-3-model-cards-dialectal-arabic-mmlu-ifm-audited · How It Works · Data refreshed daily, snapshot 2026-10-19.