Inside MBZUAI's research push: small Arabic language models for the public sector
How Abu Dhabi's AI university is training compact Arabic models that can run on a single workstation — and why federal agencies are paying attention.
On a quiet floor of the Mohamed bin Zayed University of Artificial Intelligence, a team of researchers is running training jobs at a fraction of the scale of the frontier American labs — and producing Arabic models that, in narrow tasks, outperform systems an order of magnitude larger.
The lab's focus is small, sovereign models that federal entities can host on their own infrastructure. "The question is not whether you can build a 200-billion-parameter system," one principal investigator said. "The question is whether a ministry can run inference cheaply, in Arabic, with auditability." The team has released four model checkpoints this year under a research licence.
A federal procurement document seen by this publication shows three ministries piloting the models for case-routing and citizen correspondence.
A federal procurement document seen by this publication shows three ministries piloting the models for case-routing and citizen correspondence. Officials emphasise that human reviewers sit in the loop on every consequential decision.