The Arabic-AI race: four Gulf models, one language

The Gulf has produced four major Arabic-native large language models, each backed by a national institution — a quiet race to build AI that understands Arabic the way the region actually speaks it.

The contenders

• Jais (UAE) — from Core42, with versions trained on one of the largest Arabic-first datasets assembled, spanning Modern Standard Arabic and many dialects.
• Falcon-H1 Arabic (UAE) — from TII, currently leading the Open Arabic LLM Leaderboard.
• ALLaM (Saudi Arabia) — developed by SDAIA and integrated with HUMAIN's infrastructure.
• Fanar (Qatar) — built by QCRI and benchmarked by testers from across the Arab world.

Arabic is not a translation problem — its morphology, dialects and right-to-left script need models trained for it, not bolted on after English. Four serious, well-funded efforts mean teams building Arabic-first products now have real, competitive choices.