Orchestra or an open model you run yourself?
Both run on your servers and keep data inside your network. The difference is the Arabic, the amount of hardware, what you build around the model yourself, and who is accountable when the system has to work.
What they share
- Both run on your hardware, inside your network, air-gapped when you need it.
- Neither comes with a token bill.
- Neither depends on a provider abroad to keep running.
Where they differ
- Arabic: most open models are multilingual, and Arabic is one language among many in them. The ones we tested read formal Arabic well and stumble when a customer writes in dialect. We built Orchestra from scratch for Arabic, for formal Arabic and for the dialects as customers write them. We are measuring its dialect quality now and will publish the results; the general benchmark results are on the Orchestra results page.
- Hardware: for 500 users our sizing for Orchestra is two NVIDIA A100 80 GB GPUs, while a dense open model of similar size needs three to four, and the largest open models need eight H100s. These are planning estimates until we publish a same-hardware benchmark.
- What you get: an open model is a weights file. Search over your documents that respects access rights, the safety layer, the audit log, the connections to your systems and the evaluation are all for your team to build. Orchestra arrives with those layers, and our engineers deploy it inside your environment.
- Who is accountable: with an open model your team owns everything, and support comes from the community. With Orchestra, pass criteria are agreed before you buy, the system runs in parallel before cut-over, and our engineers stay with you after launch.
- Cost: an open model has no licence fee, but it costs the time of an engineering team to build and run it, and more hardware. Orchestra has a fixed annual licence and needs less hardware. How Orchestra pricing works.
- Control of the weights: with an open model you hold the weights and can change them as you like. Orchestra is not sold as a weights file, although all of it runs on your hardware.
When an open model is the better choice
- You have a machine learning team that wants to own the model and fine-tune it itself.
- Your work is mostly in English, or in formal Arabic only.
- You are trying out an idea and do not yet need a system the business depends on.
And when Orchestra fits better
- Your customers and staff write in their dialects, not only in formal Arabic.
- The hardware budget is tight, or GPUs are hard to get.
- You have no team to build and run the retrieval, safety and logging layers.
- Your regulator asks for an audit trail and for a vendor who is accountable.
How to decide
Do not rely on our word, or on published scores alone. Test both on your own data and in your customers' dialects: a six-step method to evaluate any Arabic model before you buy. We can run that comparison with you in a pilot on your documents.
Comparing with a cloud model instead? Read Orchestra or a cloud LLM API.
Frequently asked questions
Is Orchestra open source?
No. Orchestra is a model NoorAI built; it is licensed annually and deployed by our engineers inside your environment. All of it runs on your hardware, but it is not sold as a weights file.
Is Orchestra built on top of an open model?
No. It is a 14-billion-parameter model we built from scratch at NoorAI.
Can we test Orchestra against an open model we already use?
Yes. In the pilot we agree the questions and the pass criteria with you in advance, and compare the answers on your own documents.