Four Gulf states are each building their own Arabic model

There are now four significant Arabic language model programmes running in the Gulf, and three of the four are state-funded national projects. Abu Dhabi's Technology Innovation Institute builds Falcon, whose Falcon-H1 Arabic release in January 2026 tops the Open Arabic LLM Leaderboard. Core42, part of G42 and also in Abu Dhabi, builds Jais, whose 30B model was trained on 126 billion Arabic and 251 billion English tokens. Saudi Arabia's SDAIA builds ALLaM , described as the national model, developed with more than 400 experts and 160 government bodies and trained on over three trillion tokens; ALLaM 34B powers HUMAIN Chat. Qatar's QCRI at Hamad Bin Khalifa University builds Fanar , whose Fanar-2-27B-Instruct was released around March 2026 under an Apache 2.0 licence. The strongest argument for four is sovereignty. A model a government depends on for public services, defence-adjacent analysis and citizen-facing systems is infrastructure, and states generally prefer to own infrastructure rather than rent it from a neighbour. Each programme also trains people, and the engineers who come out of them are the more durable output. There is a genuine data argument alongside it. Arabic is not one language in practice. A model built on Saudi government corpora will handle Saudi administrative Arabic better than one built in Doha, and vice versa. A good deal of the divergence between these four is real specialisation. The shared constraint is people. Four teams are solving related tokenisation problems, curating overlapping corpora, running comparable evaluations, and recruiting from the same small pool of Arabic-capable machine learning researchers — a pool that American labs also recruit from. That is the input most likely to determine how fast all four move, and it is the argument for the collaboration that already exists between them at the research level, in shared benchmarks and published evaluations. Licensing is where the practical difference may be settled. Fanar's Apache 2.0 release is the most commercially consequential decision any of the four has made, and it has attracted the least attention. An Apache-licensed model can be taken by a company in Cairo, fine-tuned, deployed commercially and shipped inside a product without asking anyone. A model available only through a national platform serves a different purpose. Over five years, the model that accumulates the most downstream deployments is the one that shapes how Arabic gets processed, and openness contributes more to that than leaderboard position does. Falcon has historically released open weights as well, which is a large part of why it has the mindshare it does. The programmes weighing how much to open are making a choice between direct control of a national platform and reach through an ecosystem — and both are legitimate objectives for a state programme.