ALLaM: Saudi Arabia's Arabic LLM — What Businesses Should Know

ALLaM is a family of Arabic-first large language models developed in Saudi Arabia by the Saudi Data and AI Authority (SDAIA) and advanced by HUMAIN, the PIF-owned AI company. Trained on Arabic and English, it powers HUMAIN Chat and is available as open weights and through cloud platforms, giving Saudi businesses a sovereign option for Arabic AI.

What is ALLaM?

ALLaM (Arabic Large Language Model) was built by SDAIA's National Center for Artificial Intelligence to handle Arabic language and culture natively rather than as an afterthought. The research team published the paper "ALLaM: Large Language Models for Arabic and English" in July 2024, reporting that expanding the vocabulary and pre-training on mixed Arabic and English data let the models learn Arabic without losing English ability, with state-of-the-art results on Arabic benchmarks such as Arabic MMLU, ACVA and Arabic Exams.

The Public Investment Fund describes ALLaM as a multimodal Arabic large language model co-developed by HUMAIN and SDAIA. HUMAIN was launched by PIF in May 2025 to build a full AI stack, from data centres to applications.

Which ALLaM models exist and where can you use them?

ALLaM is not one model but a family released over several years. These are the publicly documented milestones as of October 2026:

Date	Milestone	What it means for businesses
May 2024	IBM announced ALLaM availability on its watsonx platform (reported as the 13B version)	Enterprises can tune and deploy ALLaM with watsonx governance tooling
July 2024	ALLaM research paper published	Transparent description of training approach and benchmark results
September 2024	ALLaM 7B announced as generally available on Microsoft Azure at the Global AI Summit; SDAIA reported 7B, 13B and 70B versions	Access through a major hyperscaler's model catalogue
Open weights	ALLaM-7B-Instruct-preview on Hugging Face, now published under HUMAIN's organisation	Can be downloaded and run privately, on-premise or in an in-Kingdom cloud
August 2025	HUMAIN launched HUMAIN Chat, powered by a new ALLaM 34B model trained from scratch	A consumer and business-facing Arabic assistant hosted end-to-end inside the Kingdom
August 2026	AWS announced that HUMAIN's ALLaM will soon be available through Amazon Bedrock	Managed API access alongside other foundation models

The open 7B model card states it was trained from scratch on about 4 trillion English tokens followed by 1.2 trillion mixed Arabic and English tokens, has a 4,096-token context window and is intended as one component of a larger system, with developers expected to add their own safety measures. At launch, HUMAIN Chat was available in Saudi Arabia only, with web search, Arabic voice input for several dialects and switching between Arabic and English in the same conversation.

How does ALLaM compare with other Arabic-capable LLMs?

The table lists only published facts from each developer. Benchmarks change quickly and rarely reflect your own documents, so treat any leaderboard as a starting point, not a decision.

Model	Developer	Published sizes / details	Weights and licence	Arabic focus
ALLaM	SDAIA (NCAI) and HUMAIN, Saudi Arabia	7B (open preview, 4,096 context); 13B and 70B reported; 34B powers HUMAIN Chat	7B open on Hugging Face (Apache-2.0 tag with separate licence file); others via platforms	Arabic-first, bilingual Arabic/English
Jais 2	Inception (G42), Cerebras and MBZUAI, UAE	70B, released December 2025	Open weights on Hugging Face	Arabic-first, bilingual
Fanar	QCRI, Hamad Bin Khalifa University, Qatar	Fanar-1-9B (built on Gemma 2 9B, 4,096 context); Fanar-2-27B (built on Gemma 3 27B, 32,768 context)	Apache 2.0	Arabic-centric, bilingual
Qwen3	Alibaba, China	0.6B to 32B dense; 30B and 235B mixture-of-experts; released April 2025	Dense models Apache 2.0	Multilingual: 119 languages and dialects incl. Standard Arabic, Najdi, Egyptian and Levantine
Llama 4 Scout	Meta, USA	17B active / 109B total parameters; released April 2025	Llama 4 Community Licence	Multilingual: Arabic is one of 12 supported languages

The practical difference is focus. ALLaM, Jais and Fanar were designed around Arabic and regional culture; Qwen and Llama are global models with Arabic as one language among many. ALLaM's added advantage for Saudi organisations is provenance: a national model with in-Kingdom hosting options, which simplifies conversations with compliance teams.

What can Saudi businesses use ALLaM for?

• Arabic customer service. Chatbots and agent assistants that understand Saudi customers in Arabic and English. See our Arabic chatbot guide.
• Internal knowledge search. Retrieval-augmented generation (RAG) over policies, contracts and procedures, answering staff questions with citations to source documents.
• Document drafting and summarisation. First drafts of Arabic letters, reports, tender responses and meeting minutes, reviewed by a human.
• Document processing. Extracting fields from Arabic and English forms, invoices and correspondence, often combined with OCR.
• Government-facing communication. Formal Arabic that matches the register expected by Saudi authorities.

How do you deploy ALLaM privately?

For many Saudi organisations the main question is not which model is best but where the data goes. There are three deployment routes:

1. Managed platform. Use ALLaM through watsonx, Azure or, when available, Amazon Bedrock. Fastest to start; check which region processes your prompts and whether that satisfies your data classification.
2. In-Kingdom cloud, your tenancy. Run the open weights on GPU instances in a Saudi cloud region under your own account, network and keys.
3. On-premise. Run the model on your own GPU servers with an inference server, a vector database for RAG, role-based access control and full logging. No prompt or document leaves your network.

As a sizing rule of thumb, a 7-billion-parameter model stored at 16-bit precision needs roughly 14 GB of GPU memory for the weights alone, plus headroom for context and concurrent users; quantisation reduces this at some cost to quality. Larger models scale accordingly.

Whichever route you choose, design for the PDPL and NCA controls from the start: classify data before it reaches the model, restrict retrieval by user role, log prompts and outputs, and keep personal data in the Kingdom. In one engagement for a KSA government entity, Amwaj delivered fully on-premise AI inference with internal data pipelines and RBAC, achieving 100% API independence, full data residency and predictable cost at enterprise volume. Read the private AI case study and our guide to running Arabic LLMs on-premise.

How should you evaluate ALLaM for your use case?

1. Build your own test set. Collect 100 to 300 real questions and documents from your business, in Modern Standard Arabic, Saudi dialect and English.
2. Compare at least three models. For example ALLaM, Jais and a multilingual model such as Qwen, using the same prompts and RAG pipeline.
3. Score what matters. Accuracy against source documents, Arabic fluency and register, dialect understanding, refusal behaviour and hallucination rate, judged by native-speaker reviewers.
4. Measure operations. Latency, throughput per GPU, cost per thousand requests and context length limits.
5. Review the licence. Read each model's licence file for commercial use, attribution and acceptable-use terms.
6. Test safety. Probe for prompt injection, data leakage between users and culturally inappropriate output before going live.

Expect the answer to differ by task. A smaller model with good retrieval often beats a larger model without it, and the best choice for formal government Arabic may not be the best for informal customer chat.

What are ALLaM's limitations?

A balanced decision needs the caveats as well as the strengths:

• The open model is a preview. The downloadable 7B checkpoint is labelled "preview" and its card asks developers to add their own safety evaluation before deployment.
• Short context on the open model. A 4,096-token window limits how much text fits in one request, so long contracts need chunking and retrieval rather than pasting whole documents.
• Not every version is downloadable. The larger models are reached through platforms, which means checking where those platforms process data.
• Fast-moving availability. Platform listings change; confirm current availability in the region you need before committing.

What does it cost to run ALLaM?

There is no single price. On managed platforms you typically pay per token or per hosted instance, at rates set by each provider. Self-hosting shifts cost to GPU servers or GPU cloud instances, storage, networking and the engineers who operate them. The main cost drivers are model size, the number of concurrent users, average prompt and answer length, uptime requirements and whether you need separate environments for testing and production. For steady, high-volume internal use, owning the infrastructure often makes cost more predictable; for low or spiky volume, a managed platform is usually cheaper to start. We discuss the trade-offs in data sovereignty and on-premise AI in the GCC.

Next step

ALLaM gives Saudi organisations a credible, locally developed Arabic model with private deployment options. The value comes from the architecture around it: secure retrieval, access control, monitoring and evaluation on your own data. Amwaj Solutions builds private Arabic AI assistants on sovereign infrastructure. Explore our private AI chatbot services or contact us to plan a pilot.