Saudi AI Models: A Practical Guide to Arabic-First Language Models

For years, Arabic was an afterthought in the AI world. Global chatbots handled it passably, but they stumbled on dialects, religious context, and the cultural references that Arabic speakers take for granted. Saudi developers decided to fix that by building models that start from Arabic rather than translate into it. This guide walks through the models that matter, what each one does well, and where you can actually use them.

The Models at a Glance

ALLaM 7B Instruct

This is the model that put Saudi AI on the map for developers. Its weights are published on Hugging Face, so researchers and builders can download it, inspect it, and fine-tune it for their own needs.

Features

• Open weights released under a research license
• Instruction-tuned for conversation and task following
• Small enough to run on modest hardware
• Backed by a peer-reviewed paper presented at ICLR 2025

Strengths

• Strong results among Arabic-capable open models of its size
• Ideal for Arabic NLP research and experimentation
• Easy to adapt for custom chatbots, classifiers, and search tools
• Low cost to test before committing to anything larger

ALLaM 13B Instruct

The 13B tier is aimed at businesses that want a managed service instead of hosting a model themselves. It sits in IBM's watsonx catalog alongside other foundation models, so teams already working in that environment can add it with little friction.

Features

• Hosted as a managed endpoint with pay-per-token pricing
• Handles both Arabic and English prompts
• Suited to enterprise workflows such as summarisation, drafting, and question answering

Strengths

• More headroom than the 7B version for longer, more nuanced responses
• No infrastructure to maintain
• Good fit for companies already using IBM's tooling

ALLaM 2 7B Instruct

The second-generation 7B model refines the original and is available through Microsoft's Azure AI Foundry. It gives Azure users a way to choose an Arabic-tuned option from the same catalog as the big international models.

Features

• Improved instruction following over the first release
• Deployable as a managed endpoint with usage-based billing
• Sits beside other major models, which makes side-by-side testing simple

Strengths

• Easy to benchmark against general-purpose models on real Arabic tasks
• Convenient for teams already standardised on Azure
• Compact enough to keep per-request costs reasonable

ALLaM 34B

This is the flagship. Trained from scratch as an Arabic-centric foundation model, it is the largest in the family and the one HUMAIN highlights most. It powers HUMAIN Chat, a conversational app on web, iOS, and Android, and it is the only way to use this particular model, since there is no public developer API for it.

Features

• Built for Arabic speakers and for Islamic and Middle Eastern cultural context
• Hosted entirely on local infrastructure, aligned with the Personal Data Protection Law
• Delivered through a free consumer chat app
• Independently benchmarked by Cohere as a leading Arabic-language model in the Arab world

Strengths

• Natural fluency in Modern Standard Arabic and regional expression
• Better cultural and religious sensitivity than most global chatbots
• Strong data residency story for privacy-conscious users
• The most capable option in the family for open-ended conversation

Metabrain

Aramco built Metabrain for its own workforce rather than for the public. It is trained on decades of the company's technical and operational data, which makes it a good example of a domain-specific model rather than a general chatbot. Public reports give different parameter counts, and the company has also spoken about developing a much larger model, so treat the exact size as a moving target.

Features

• Trained on internal energy-sector knowledge
• Used by employees for training and day-to-day work
• Part of a wider push into AI for operations such as reservoir simulation and gas flaring monitoring

Strengths

• Deep command of industry-specific terminology and processes
• Keeps sensitive corporate data inside the company
• Shows how a private model can outperform general tools on narrow, specialised tasks

Noon

Noon comes from Naseej, a technology company with a strong footprint in education and libraries. It was positioned as one of the early Arabic models in the region and has been used to power practical tools instead of a standalone chat product.

Features

• Powers applications such as library cataloguing agents, course evaluation tools, and exam creation
• Focused on Arabic-language relevance
• Offered through business and institutional solutions

Strengths

• Proven in real education workflows
• Saves time on repetitive academic and administrative tasks
• A good reference for how Arabic models can be packaged as ready-to-use products

Choosing the Right Model

Your choice depends on what you want to do. Researchers and developers who like to tinker will get the most from the open ALLaM 7B. Companies that need a managed service will find the 13B on IBM or the ALLaM 2 on Azure more convenient. If you simply want to chat in Arabic with the strongest model available, HUMAIN Chat running ALLaM 34B is the place to go. Organisations with their own data and a narrow use case can look at the Metabrain approach, where a model is trained on proprietary knowledge and kept in-house.

One caution applies across the board. Parameter counts, hosting options, and pricing change quickly in this space, so confirm the current details with the provider before building anything around a specific model.

› More Article