Saudi Arabia's national model involved 160 government bodies

ALLaM, Saudi Arabia's national Arabic language model, was built by SDAIA — the Saudi Data and Artificial Intelligence Authority — with the involvement of more than 400 experts and 160 government bodies, and trained on more than three trillion tokens. Its 34-billion-parameter version powers HUMAIN Chat. The number that stands out in that description is 160. Not the parameters, not the tokens, but the count of government entities coordinated to build it. What that coordination buys is data. A language model's quality is bounded by its corpus, and the corpus that matters for a national model is the one no one else can assemble: administrative Arabic, legal text, ministerial correspondence, standards documents, regulatory guidance, the accumulated written output of a state. None of that is on the open web. A lab scraping Common Crawl cannot obtain it. A model with 160 government bodies contributing can, and that is a durable advantage in exactly the applications a government cares about — drafting, summarising, answering citizen queries about entitlements, routing requests between agencies. It is also a substantial piece of institutional work in its own right. Getting 160 entities to agree data-sharing terms, classification and provenance is the kind of groundwork that outlasts any single model, because the agreements and the data governance survive into whatever is trained next. The trade-off is cadence. A programme built through multi-party data agreements moves on the timetable of that coordination, where frontier model development runs on a shorter cycle. Both timetables are doing different work, and the corpus that results is not something a faster process would have produced. There is also a design consequence worth naming. When one authority builds the national model, sets the national AI framework and coordinates national data governance, the result is a consistent standard for how Arabic is handled in official use — which is the point of a national model, and a decision whose effects will outlast the current generation. Three trillion tokens is a substantial training run and not an extraordinary one by 2026 standards; frontier labs passed that scale some time ago. ALLaM's distinguishing feature is composition rather than volume — a much higher Arabic share than any general model, and a slice of Arabic that is not otherwise available. Which is the right design for a national model. Competing with frontier labs on raw scale is not the available game for a state programme; competing on data nobody else holds is.