Fanar-1-9B-Instruct: benchmark results
QCRI's Fanar-1-9B-Instruct, an Arabic-English instruction-tuned model continually pretrained from Gemma 2 9B on 1T tokens, covering Modern Standard Arabic and Gulf, Levantine and Egyptian dialects. Provider: QCRI. Released 2025-06-01. Access: Open.
Aggregate Rating 1541 ± 16, rank #635 of 1635 rated models, from 202 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ArabicDialectSafety - Harm Category Classification | 69 | Macro-F1 (%; which of seven harm categories a test prompt be | 100 |
| QIMMA - ArabLegalQA | 73.08 | BERTScore F1 (%) | 97.3 |
| Open Arabic LLM - Aratrust MentalHealth | 96.05 | Accuracy (%) | 97.2 |
| Open Arabic LLM - Arabic MMLU Arabic Language (Grammar) | 69.04 | Accuracy (%) | 96.6 |
| Open Arabic LLM - Madinah QA Arabic Language (Grammar) | 68.77 | Accuracy (%) | 95.1 |
| QIMMA - MedAraBench | 48.53 | Accuracy (%, log-likelihood multiple choice) | 92.7 |
| Open Arabic LLM - Alghafa Multiple Choice Facts Truefalse Balanced Task | 94.67 | Accuracy (%) | 92.3 |
| Open Arabic LLM - Madinah QA Average | 73.35 | Accuracy (%) | 91.4 |
| Open Arabic LLM - Arabic MMLU Arabic Language (High School) | 45.9 | Accuracy (%) | 90.7 |
| Open Arabic LLM - Aratrust PhysicalHealth | 91.78 | Accuracy (%) | 90.7 |
| QIMMA - AraTrust | 88.51 | Accuracy (%, log-likelihood multiple choice) | 89 |
| Open Arabic LLM - Aratrust Offensive | 92.75 | Accuracy (%) | 88.2 |
Interactive version: theaggregate.ai/model?slug=fanar-1-9b-instruct · How It Works · Data refreshed daily, snapshot 2026-11-05.