ALLaM explained: Saudi Arabia's national Arabic LLM, and where it fits
What ALLaM is, how SDAIA built it, where you can actually run it, how it compares with Jais, Falcon Arabic and the frontier models — and how to decide whether it belongs in your Saudi deployment.
ALLaM is the Kingdom's national Arabic large language model, developed by the National Center for Artificial Intelligence at SDAIA. It matters for two separate reasons, and they are worth keeping apart. As a model, it is a credible Arabic option you can actually deploy. As a signal, it tells you that Arabic capability is being treated as national infrastructure rather than as a feature other people's models might get around to.
What ALLaM actually is
It is a generative model built to answer questions across knowledge domains in Arabic, drawing on trusted sources from the Kingdom and the wider Arab world, with English content included so responses are not stranded in one language. SDAIA describes mobilising a large number of public entities to assemble the underlying corpus — one of the largest Arabic training datasets built to date. That corpus, more than any architectural choice, is the thing that is genuinely hard to replicate.
The published family spans several sizes, including variants initialised from Llama 2 weights and a smaller model trained from scratch, with instruction-tuned versions released for practical use. Reported results have placed it at or near the top of Arabic benchmarks in its size class. Treat any specific benchmark claim, including that one, as a starting point rather than a verdict — you still have to measure it on your own data.
Where you can run it
| Route | What it gives you | Worth checking |
|---|---|---|
| IBM watsonx | Announced at IBM Think in May 2024; studio access to train, tune and deploy with enterprise governance tooling. | Which variant is currently served, and the governance features your review actually needs. |
| Microsoft Azure | Made available to Azure customers in the Kingdom and globally; the model was itself developed and trained on Azure. | Region availability for your tenancy, and whether inference can stay in-Kingdom. |
| Open weights | Model weights published for download, enabling self-hosting and fine-tuning on your own infrastructure. | Licence terms — distribution has combined an SDAIA licence with the upstream Llama 2 community licence. |
The licensing point deserves emphasis. A model derived from Llama 2 weights inherits obligations from that licence alongside SDAIA's own terms. If you intend to fine-tune and deploy commercially, read both before your legal team does.
ALLaM against the alternatives
| Option | Strongest case for it | Where it gives ground |
|---|---|---|
| ALLaM | Arabic-first corpus, national provenance, deployable through major clouds and self-hostable. | Reasoning and tool-calling maturity against frontier models; verify on your own tasks. |
| Jais / Falcon Arabic / Fanar | Serious regional Arabic efforts with their own corpora and research pedigree. | Ecosystem tooling and hosting options vary; check what is deployable where. |
| Frontier general models | Best-in-class reasoning, instruction following and tool calling, strong Modern Standard Arabic. | Residency and sovereignty are constrained by where the provider operates. |
| Open multilingual weights | Full control over hosting and fine-tuning, fixed cost at volume. | You own serving, latency engineering and evaluation. |
How to decide, for a voice deployment
In a voice system the language model sits between transcription and synthesis, which changes what matters about it. Four tests decide the question, in this order.
- Dialect comprehension — feed it transcripts of real Najdi and Hijazi speech, including the errors your speech-to-text actually makes. Robustness to imperfect input matters more than eloquence.
- Tool calling — measure how often it calls the right function with correctly typed arguments across a hundred varied turns. This separates production-ready from impressive.
- Time to first token — measured with your real system prompt and conversation history, because it becomes part of the silence the caller hears.
- Instruction adherence in Arabic — run your actual prompt and check it respects scope, length and register rather than drifting.
Our general advice holds here: keep the model layer swappable behind a thin interface and re-benchmark on a schedule. A national model is a strong option and a strong signal, but it should still earn its place on your evidence rather than on its passport.
Why a national Arabic model exists at all
Arabic is spoken by hundreds of millions of people and has been chronically under-served by models trained predominantly on English text. SDAIA frames ALLaM as part of serving and empowering the language — preserving its integrity while enriching Arabic content across technical, cultural, literary and scientific domains — and ties that to Vision 2030's ambition for the Kingdom to lead in advanced technologies. Whatever you conclude about the model, the direction is unambiguous: Arabic capability is being built here, deliberately, as national capability.
What Voho is building
Voho is developing its own Arabic speech and language models tuned specifically for Saudi dialects and for the conditions real calls happen in — narrowband telephony audio, code-switching, and the names, numbers and dates that decide whether a call succeeds. We are not going to publish parameter counts or benchmark claims before they can be independently checked. What we will say is how we work: models are components with measured jobs, evaluated against real calls from the deployments they serve, and replaced when the evidence says something else is better. That applies to our own models too.
Sources
Frequently asked
- What is ALLaM?
- ALLaM is Saudi Arabia's national Arabic large language model, developed by the National Center for Artificial Intelligence at the Saudi Data & AI Authority (SDAIA). It is a generative model built to answer questions across knowledge domains in Arabic, trained on one of the largest Arabic corpora assembled to date with English content included alongside it.
- Where can I use the ALLaM model?
- ALLaM has been made available through IBM watsonx, announced at IBM Think in May 2024, and to Microsoft Azure customers in the Kingdom and globally. Model weights have also been published for self-hosting and fine-tuning. Check the current licence terms before commercial use, since distribution has combined an SDAIA licence with the upstream Llama 2 community licence.
- Is ALLaM better than GPT or Claude for Arabic?
- It depends entirely on the task. ALLaM's advantage is an Arabic-first corpus and national provenance; frontier general models typically lead on reasoning, instruction following and tool calling. For a voice deployment the decision should come from measuring dialect comprehension, tool-calling accuracy, time to first token and instruction adherence on your own transcripts rather than from published benchmarks.
- Can ALLaM be fine-tuned on my own data?
- Yes — weights have been published for self-hosting, and platform routes such as watsonx offer tuning and deployment tooling. The practical constraints are licence compliance, having enough correctly labelled Arabic data from your own domain, and holding back an evaluation set so you can prove the tuned model is actually better.
Keep reading