Seraj: an Arabic-first AI model built for institutions

Inception42 and Microsoft launched Seraj on 2 July, an enterprise AI model built on GPT-4.1 and mid-trained on curated Arabic datasets covering linguistics, cultural knowledge and domain content. The stated target list is telling: government services, legal analysis, education, media, financial services — and Islamic studies, named explicitly as an enterprise workload. Frontier models have consistently performed better in English than in Arabic, which quietly taxes every Arabic-speaking institution that adopts them; Seraj is one of several serious attempts (alongside Falcon-H1 Arabic and Saudi Arabia's Allam) to close that gap. The approach matters too: rather than training from scratch, the team adapted a frontier model with targeted Arabic mid-training, keeping its reasoning while fixing its blind spots. Whether communities end up building on such models or merely renting them is the open question — but the existence of Arabic-first infrastructure is the precondition for either.

This is a QeRN summary by Ahmed Qerni. Read the original at Gulf Business: https://gulfbusiness.com/en/2026/artificial-intelligence/inception42-launches-arabic-ai-model-with-microsoft/.