Orchestra: an Arabic language model that runs on your servers.
A 14-billion-parameter model we built from scratch at NoorAI. It runs on your own GPU server inside your network, for one fixed annual licence instead of a bill per token, air-gapped when you need it.
What it is
Orchestra is NoorAI's main model. Arabic is its first language, with English for mixed environments. We built it for formal Arabic and for the Arabic customers actually write, in Egyptian, Gulf and Levantine. It streams its answers and is built to run agents, not only to chat. We are now measuring how it does on dialects against the available models, and we'll publish the results when the measurement is complete. New to the idea? Read what an on-premises LLM is, and when to choose one.
Where it runs
Your data centre, or an in-country private cloud
Why we built our own instead of deploying an open model
- Dialects: the models we tested read formal Arabic well and stumble when a customer writes in dialect, both in understanding it and in writing it. We couldn't find a light, strong Arabic model built from scratch to run on-premises, so we built Orchestra.
- A fraction of the infrastructure: open models of similar quality need serious on-premises hardware. Orchestra runs on one GPU-class node, not a cluster.
- Concurrency: the model and its serving stack are optimised together, so the same hardware serves more users at once.
Why on your servers, for a fixed price
- Data doesn't leave: in Egypt and across the Gulf, data protection rules restrict moving customer data abroad, and some sectors are not allowed on foreign cloud at all. When the model, the index and the logs run on your hardware, there is no transfer to begin with. The rules, country by country.
- Nobody abroad can switch it off: a model hosted by a foreign provider can be cut off by a political decision or a change in the provider's policy. If your business runs on it, you are hostage to that decision. Orchestra depends on no outside service to run.
- A budget you know in advance: nobody can predict how many tokens an organisation will consume in a year. So Orchestra comes with a fixed annual licence: no token meter, no bill that grows with use.
What comes with it
- Long context for contracts, filings and multi-hundred-page regulatory submissions.
- Citation-grounded retrieval with Arabic-capable OCR: every answer shows its sources.
- Layered safety: a dedicated classifier plus policy rails your regulator can read; every decision logged.
- Arabic speech-to-text and text-to-speech, also on-premises, with versions trained for specific sectors.
How you get it
Orchestra is not sold as a weights file. NoorAI's engineers deploy it inside your data centre, or in a private cloud or dedicated VPC in your country, together with the retrieval, safety and orchestration layers. We deliver in three phases: Discover, Deploy, Operate. Every deployment has pass criteria agreed up front and a parallel run before cut-over.
The licence is annual, at a fixed price per deployment, and the price depends on the deployment size we agree together in Discover. Numbers on dialect quality and throughput we publish once measured, not before. Until then, ask for a live demonstration on your own documents.
Frequently asked questions
Does Orchestra run fully air-gapped?
Yes. The model, retrieval, safety and logs all run inside your premises, and nothing needs a connection to an outside service to work. It can also be deployed in a private cloud or a dedicated VPC in your country.
What hardware does it need?
One GPU-class node, not a cluster. 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.
Which Arabic dialects does it understand?
We built it for formal Arabic and for Egyptian, Gulf and Levantine as customers write them, with English for mixed environments. We are measuring its dialect quality against the available models now and will publish the results, so we promise no numbers before then.
How is it priced?
A fixed annual licence per deployment. No token meter and no per-query fees, so the budget is known before the year starts. The price depends on the deployment size, agreed in the Discover phase.
What if a provider abroad changes its policy or stops the service?
Nothing changes for you. Orchestra runs on your hardware and depends on no outside service to run, so it keeps working whatever a cloud provider or a foreign company decides.
Can we get the model weights?
Orchestra is not sold as a weights file. NoorAI's engineers deploy it in your environment together with the retrieval, safety and orchestration layers, and everything stays on your hardware.
Does it connect to our existing systems?
Yes, through the agent platform: ERP, CRM and databases through the Model Context Protocol, and channels such as WhatsApp, web, email and voice. Sensitive actions go through approvals that follow your organisation's delegation of authority.
How can we evaluate it before we decide?
Ask for a live demonstration on your own documents. In any project we agree the pass criteria up front and run the system in parallel with your current work before cut-over.