Nawah-VL — وصف الصور بالعربية

SigLIP2 vision tower grafted onto an Arabic language model. Pick a backbone, upload an image, get a Modern Standard Arabic caption.

الصورة

Model

Nawah-VL-50M · better (145M total) Nawah-VL-25M · smaller (118M total)

الوصف

What to expect. It names the main subject and its colour reliably and stops cleanly. It reads text inside images poorly and will occasionally state something confidently wrong — the language model is 25M/50M parameters, so coarse-but-correct is the ceiling.

	chrF++	perplexity	grounding gap
25M	27.4	12.2	2.09
50M	29.0	8.6	2.36

جرب صورة من دول

الصورة	Model