Research
Open Arabic medical AI, published with its method
We build Arabic-language models for medicine and publish them openly on Hugging Face, with their methods. The Hakeem-7B method is peer-reviewed in Knowledge-Based Systems (Elsevier, 2026).
Hakeem-7B
A 7-billion-parameter model that reasons about medicine in Arabic, trained past the 60% line on MedAraBench.
- 0.600 on MedAraBench
- Method peer-reviewed, 2026
vioBERT-v3
An Arabic medical language encoder: MARBERTv2 domain-adapted on 1.12 million Arabic medical documents.
- 85.9% fill-mask top-5
- Perplexity down 82.7%
Ar-CXR
Reads a chest X-ray and writes a four-section report in Arabic, first on every Arabic metric tested.
- 40.4 clinical Jaccard, first of three
- 11 findings at 0.789 AUROC
What these models are not for
These are research models. They are not for diagnosis, triage, dosing or treatment decisions, and their output needs verification by a qualified clinician. We keep them separate from the production systems hospitals run.
Let’s build your next system
Tell us what has to work, and we will scope it honestly. We work with hospitals, ministries and enterprises across Egypt and MENA.
Enterprise engagements begin under NDA. We reply by email.