Research · Open Arabic vision–language model
Ar-CXR: the first native Arabic chest-X-ray vision–language model
No open model could read a chest X-ray and write a report in Arabic. We coupled a frozen RAD-DINO vision encoder to a Falcon-H1-7B Arabic decoder (7.4 billion parameters, only about 4% trained) that grounds 11 findings and writes a four-section report, ranking first on every Arabic metric.
- Architecture
- Frozen RAD-DINO vision encoder + Falcon-H1-7B Arabic decoder
- Size
- 7.4 billion parameters, about 4% trainable
- Training images
- 223,462 CheXpert images
- Findings
- 11, at 0.789 macro-AUROC
- Report
- Four sections, in Arabic
- Evaluation
- Identical 200-image, image-only Arabic protocol
Clinical Jaccard on Arabic report generation, same prompt across all models, 2026.
Where it stands
Ar-CXR grounds 11 findings at 0.789 macro-AUROC and leads METEOR, chrF, BERTScore and CIDEr, all under an identical 200-image, image-only Arabic protocol.
Limits
This is a research model, published openly with its method. It is not for diagnosis, triage, dosing or treatment decisions, and its output needs verification by a qualified clinician. We keep research models separate from the production systems hospitals run.
Updated