ALLaM

The open model is small, trained from scratch, and still a preview

ALLaM-7B-Instruct-preview appeared on Hugging Face in February 2025, and its page was last updated in July 2025. The card says it was trained from scratch rather than adapted from a foreign model: 4 trillion English tokens first, then 1.2 trillion mixed Arabic and English tokens. “Preview” means what it says, an early release rather than a finished product, and no newer open ALLaM model had appeared by the time of this review.

The page is tagged Apache 2.0, but the card points to a licence file we could not find in the repository, so read the terms yourself before commercial use. A 7B model runs on a single consumer GPU. There is no official GGUF build, but community members have published quantised versions you can load in LM Studio; it is not in Ollama’s official library. The steady stream of community fine-tunes built on it, many published between August and October 2026, shows that Arabic developers do use it as a base.

HUMAIN Chat runs the bigger ALLaM, behind a border

HUMAIN launched HUMAIN Chat on 25 August 2025, saying it runs on ALLaM 34B and is available on iOS, Android and the web. It launched inside Saudi Arabia only, with a promise to roll out “soon” to other Arabic-speaking countries. We found no later official announcement of that expansion, and the chat site blocks automated requests, so if you are outside the Kingdom, try the link yourself before planning around it.

On 26 August 2026, Microsoft and HUMAIN announced plans to bring ALLaM to Microsoft Foundry and Microsoft 365 Copilot. The announcement speaks of plans and “the coming months”, with no date or pricing, and we found nothing confirming the models are live there yet.

Saudi Arabic is the test that matters

HUMAIN says ALLaM understands dialects, from classical Arabic to local forms, and that it was refined with input from more than 600 domain experts and 250 evaluators. Those are the company’s claims, and we have not tested them. A better check takes five minutes: ask it, in Najdi or Hijazi dialect, to draft a leave request to your manager, then ask for the same letter in formal Modern Standard Arabic, and compare both with what ChatGPT writes. If you are trying the open 7B model, set expectations accordingly: it is good at narrow tasks such as classification, short summaries and rewriting, and weaker at long chains of reasoning.

Who it suits

The open model suits developers and researchers who want a light Arabic starting point to fine-tune on their own domain data. HUMAIN Chat suits users in Saudi Arabia who prefer a locally built assistant. We would not pick ALLaM if you need a public API today, since none has been announced with pricing, or if you want the newest open Arabic model: Jais 3 offers newer sizes under a clear licence.

For how the Arabic projects compare and when to choose each, see our guide Arabic AI Models: Does a Model That Speaks Arabic Really Understand It?, and to run the open model yourself, read How to Run AI Models on Your Own Computer. To try it: HUMAIN Chat, or the open model page.