Question 1

What does Arabic-first actually mean?

Accepted Answer

It means the Arabic is trained in rather than translated on. The pre-training corpus is built for Arabic language-model pre-training rather than scraped as an afterthought, the models are trained on it, and they are measured on Arabic evaluations we had to build ourselves because the existing ones were too narrow to settle the question. An interface localised after the product was designed is a different thing, and this site tries never to use the phrase for that.

Question 2

Which Arabic dialects do you handle?

Accepted Answer

Lahjawi is a family of dialect translation models covering dialect to dialect and dialect to Modern Standard Arabic, and Sawt Najd generates Najdi speech. Coverage differs between models, so it is stated on each model's own page rather than as one claim for the whole catalogue.

Question 3

Do you handle Islamic and classical Arabic text?

Accepted Answer

It is one of the two research tracks, not a side project. Kawn Embed Islamic retrieves across the Quran, Hadith, Fiqh and Tafseer; the Islamic QA Classifier routes jurisprudential questions and the Islamic Quotation Extractor identifies verbatim quotation; and Nabr is an acoustic model for Quranic recitation covering speech recognition and Tajweed analysis. The tracks page lists what each of them has behind it.

Question 4

Can the models run on our own infrastructure?

Accepted Answer

Most of them, and each model page states its own options rather than the site making one blanket claim. The catalogue records cloud, on-premises and sovereign deployment per model, so a model that is cloud-only is visibly cloud-only. Where on-premises is available it is the same model running inside your boundary, not a reduced edition of it.

Question 5

Where does our data live, and who owns it?

Accepted Answer

On an on-premises or dedicated deployment nothing leaves your environment unless you decide it should, and servers can be hosted inside Saudi Arabia. Data ownership and sovereign deployment are part of the architecture rather than a tier that is sold on top of it. The Trust Center holds the detail, including what residency means for each deployment shape.

Question 6

How do we get API access?

Accepted Answer

Through Kawn Console at dashboard.kawn.ai, which is the developer surface for the API. The developers page lists every model in the catalogue that records cloud access, so a model missing from that list is one you cannot call today rather than one nobody documented.

Question 7

Which models are open and which are commercial?

Accepted Answer

Every model row records its licence type and the models page filters on it. Open is not the same as published: a model can be openly licensed with no weights released yet, which is why the open source page lists only the rows that actually have something you can download or read.

Question 8

How does licensing work?

Accepted Answer

Open weights carry their licence in the Hugging Face repository they are published from, and that copy is the one that governs. Commercial models are not published as weights and access is arranged directly. No prices are published on this site, so nothing you read here is a quote.

Question 9

Can you adapt a model to our domain?

Accepted Answer

Yes, and the services page describes how: take the catalogue model closest to your task, specialise it on your material, and agree the evaluation set before any model is touched. Kawn Embed ships in three domain variants for exactly that reason.

Question 10

Do you publish your research?

Accepted Answer

Yes, and it is the same work the products run on rather than a separate lab output. Papers, open weights, released corpora and the benchmarks are all listed under Kawn Lab, each one linked to the artefact itself, so a claim on this site can be checked rather than taken.

Question 11

How do you evaluate an Arabic model?

Accepted Answer

By building the evaluation when none exists. Misraj-DocOCR, SadeedDiac-25 and Tarjama-25 were each built because the available Arabic evaluations were too narrow or too biased towards English source text to settle the question, and each is published alongside the model it measures. The benchmarks page carries the metric, the dataset and the date for every result.

Question 12

Do you build custom solutions?

Accepted Answer

Yes. We start from the specific challenge your organization faces, not a generic template, then build the right solution around it, on our own models and technology: Kawn Lab models, Baseer, Sada, and Seamless. The solution can run in the cloud, on-premises, or sovereign inside the Kingdom where that is required, and it connects to your existing systems and data instead of working apart from them. The process starts with a consultation to define the challenge precisely, then we design, build, and deliver the solution with your team, from the first scope to it running in your environment. Book a consultation to start.

Frequently asked questions

The models

Deployment and data

Access and licensing

Research and openness

Let's talk about what you're trying to build.