When a Gulf organisation looks for “local Arabic AI”, it usually means two things at once: a system that runs locally, inside the organisation or at least inside the country, and a system that understands Arabic as it is written locally, in Gulf, Najdi, Hijazi and Egyptian Arabic, not only in formal Arabic. Vendors use the word “local” loosely. This guide separates the meanings and gives you questions to check each one.

“Local” as a place: three degrees

  • Hosted in-country: the model runs in a data centre or cloud inside the country’s borders. That settles where the data sits, but the provider may still be the one who runs the model, updates it and keeps the logs.
  • Running inside your premises: the model, its serving stack, the indexes and the logs sit on hardware you control, in your data centre, a private cloud, or a dedicated VPC in your country.
  • Air-gapped: the highest degree. No outside connection at all, and updates arrive through a channel you control. We explain what that means in practice in what “air-gapped” really means.

The data decides the right degree before the technology does. Our rule of thumb: pick the most isolated option your team can run well. More in where your Arabic AI runs, and on the gap between in-country hosting and full sovereignty in sovereign AI in the Gulf.

“Local” as a language: Arabic as people write it

A system that runs inside your building but reads Arabic as if it were translated English is not local in the sense that matters to your users. Customers write in their dialects and mix Arabic with English in one message, while staff write official letters in formal Arabic. A model that handles carefully written formal Arabic can still stumble on a WhatsApp message from a customer in Riyadh or Jeddah.

Ask the vendor: was the model trained on Arabic text with an Arabic-designed tokenizer, or is it an English model with Arabic added on? Which dialects does it understand and answer in? Then do not stop at the answer: test it on your real messages, as in how to evaluate an Arabic language model before you buy it.

Why Gulf buyers ask for it now

  • Rules are moving towards keeping data at home: each country has its own, and we summarised what matters for AI in Saudi Arabia, the UAE, Qatar, Kuwait, Oman and Bahrain. These summaries are not legal advice; the decision sits with your legal team.
  • Sensitive data: government correspondence, banking customer data and health records are hard to send to an outside service, even where the rules allow it.
  • A fixed version: a system your regulator has approved should not change because a provider updated its model. Running locally means the version is yours, and you decide when to update it after testing.
  • Predictable cost: the bill does not grow with every conversation, as it does with pay-per-use. The full comparison is in on-premises or cloud language model.

Five questions that show whether a solution is really local

  1. Where does the model itself run, not just the interface? Some solutions host the application locally and send the text to a model running somewhere else.
  2. Does it work with no outside connection? Ask to see it running with the connection cut.
  3. Where are the logs and indexes kept, and who can read them?
  4. Who decides when the model is updated, and does the current version stay fixed until you have tested the new one?
  5. Does it understand your customers’ and staff’s dialects? Test it on a sample of your real messages, not the vendor’s examples.

How we build it at NoorAI

Orchestra is a 14-billion-parameter Arabic language model, trained from scratch at NoorAI on curated Arabic text with an Arabic-designed tokenizer. It understands Modern Standard Arabic and Gulf Arabic, Najdi and Hijazi included, plus Egyptian and Levantine, with English for mixed environments. It runs in your data centre, or in a private cloud or dedicated VPC in your country, air-gapped when needed, on one GPU-class node rather than a cluster.

Our headquarters and engineering team are in Cairo, and we work with partners in Dubai and Riyadh, across Saudi Arabia, the UAE, Egypt and the Gulf. We publish dialect-quality numbers once they are measured, not before. Until then, ask for a live demonstration on your own documents.

Frequently asked questions

What is the difference between local AI and sovereign AI?

“Local” describes where the system runs. “Sovereign” is wider: who owns the data and the version, who can audit the model’s behaviour, and what happens if the provider’s terms change. Running inside your premises answers most of those questions at once.

Can we run Arabic AI locally without our own data centre?

Yes. Orchestra can be deployed in a private cloud or a dedicated VPC with an in-country provider, with the same security controls.

Is local AI lower quality than cloud services?

Not necessarily, and the answer differs from task to task. The only reliable way is to compare the options on your data and your users’ dialects, with pass criteria written before you see the results.

What hardware do we need?

Orchestra runs on one GPU-class node. The exact sizing depends on your number of users and the volume of documents, and we size it with you in the Discover phase, before anything is bought.

Related pages