HUMAIN, the Saudi AI company, said on 11 November 2026 that a 3 trillion parameter open-weight model will be available in Saudi Arabia in December, with its complete AI stack due on 15 December. Only the size, the open-weight description and the December timing have been stated so far; no model name, language coverage, compute figures or licence terms have been published.
What HUMAIN has announced
Arab News reported (12 November, updated 13 November) that HUMAIN chief executive Tareq Amin spoke on Wednesday evening at the HUMAIN+ announcement at HUMAIN’s offices in Majdoul Tower, Riyadh. The report summarises the plan as the “world’s largest open-weight model, 3 trillion parameters” becoming available in Saudi Arabia in December. Amin said the complete stack arrives on “December 15” and that the model and infrastructure would “come together in December”.
Amin described the model as using “88 percent less cost and token consumptions than available models outside of the country”. Arab News does not say what that comparison is measured against. Middle East AI News (13 November) describes the model as still unnamed, notes that the 88% figure is a company assertion with no independent benchmark published, and reports that HUMAIN said development took about 15 months.
What has not been stated
- The model’s name, architecture and training data.
- Which languages it supports, including how well it handles Arabic.
- The compute needed to run it, and whether the weights will be downloadable outside Saudi Arabia.
- The licence under which the weights would be released.
An open weights model publishes its trained parameters, but that term alone says nothing about the training data or the licence. Until HUMAIN publishes a model card or release notes, the open-weight claim is a description, not a documented property, and the size is a headline figure for a large language model that nobody outside HUMAIN has been able to test.
Why it matters for research access
This section is editorial interpretation, not reporting. Researchers who study or use national-scale language models need three things before a release is useful: the weights or an API they are allowed to use, documentation that supports citation and reproduction, and a way to run it at a realistic cost. A size announcement settles none of them. The phrase “available in Saudi Arabia” leaves open whether access is limited to in-country users or infrastructure.
For Arabic-language research, the announcement gives no language details, so any claim about Arabic performance would be speculation. Research offices evaluating the model for projects should wait for published evaluations and the licence text, and record the exact model version used, in line with their generative-AI disclosure practice.
The announcement also fits a broader pattern of governments and national companies building their own model capacity, which CASRAI has covered in EU leaders’ Dublin AI strategy discussion and in the frontier LLM landscape overview.
What to watch next
- 15 December 2026: the date HUMAIN gave for its complete stack.
- Publication of the model name, weights location, licence and any independent benchmarks.
- HUMAIN+, an AI education and training platform that Arab News says is planned to launch operationally in the first quarter of 2027.
