ALLaM — Saudi Arabia's National Arabic Language Model

ALLaM represents something unprecedented in the history of artificial intelligence: a large language model developed as an explicit instrument of national strategy, backed by sovereign data access that no private company could replicate, and positioned as the linguistic foundation for an entire nation’s digital transformation. Built by the National Centre for Artificial Intelligence operating under Saudi Arabia’s Data and Artificial Intelligence Authority, and now managed by the kingdom’s purpose-built national AI company HUMAIN, ALLaM carries the weight of Vision 2030’s technological ambitions in a way that no other Arabic language model can claim.

The model’s name — an Arabic word meaning “knowledgeable” or “scholarly” — signals its intended role: not merely as a chatbot or text generator, but as a knowledge system capable of understanding and reasoning about Arabic at a depth that matches human expertise. The development program mobilized 16 Saudi government entities to aggregate training data, engaged 400 subject matter experts to query and test the model through more than one million prompts, and assembled the world’s largest Arabic language training dataset at 500 billion tokens. This institutional mobilization reflects a fundamental insight: building a truly capable Arabic AI requires access to data, expertise, and resources that only a sovereign entity can command.

Model Architecture and Versions

ALLaM’s development has proceeded through multiple model sizes and architectural iterations, reflecting an evolving strategy that balances research ambition with deployment pragmatism.

The initial model family was built on Meta’s Llama 2 architecture, with SDAIA’s NCAI developing 7-billion, 13-billion, and 70-billion parameter versions. ALLaM-1-13b-instruct, the first publicly available version, was pre-trained on 3 trillion tokens in Arabic and English — a massive corpus that enabled the model to develop broad knowledge across multiple domains while maintaining strong Arabic language competence.

ALLaM-2-7B represented a refinement focused on deployment efficiency. This autoregressive transformer model was designed specifically for production environments requiring advanced Arabic and English natural language understanding with manageable computational requirements. The 7-billion parameter size makes it deployable on standard enterprise GPU infrastructure, enabling organizations to run ALLaM internally without the latency and compliance concerns associated with cloud-only API access.

The transformative release came with ALLaM 34B in 2025, the first ALLaM model built entirely from scratch rather than adapted from an existing architecture. Developed under HUMAIN’s stewardship, ALLaM 34B was designed and trained specifically for Arabic language performance, with particular emphasis on Saudi Arabian dialects, cultural knowledge, and regulatory compliance. The model was fine-tuned using Saudi subject matter expertise across government, business, legal, healthcare, and educational domains, creating a foundation model that understands the specific contexts in which Saudi organizations operate.

Training Data: Sovereign Scale

The training data assembled for ALLaM represents perhaps the most significant competitive advantage of any Arabic language model. SDAIA mobilized 16 public entities to contribute data, creating a 500-billion token Arabic dataset that remains the largest purpose-built Arabic language training corpus. The dataset includes the full text of 300 Arabic-language books carefully selected for linguistic quality and domain coverage, government documents spanning decades of Saudi administrative history, academic publications from Arabic-language research institutions, news archives from major Arabic media organizations, and curated web content filtered for quality and authenticity.

The curation process was intensive. Rather than relying solely on automated quality filtering — the approach used by most language model developers — SDAIA engaged 400 subject matter experts who generated more than one million prompts to test and refine the model’s knowledge and reasoning capabilities. These experts spanned medicine, law, engineering, education, Islamic studies, Arabic linguistics, and other fields, ensuring that the model’s training captured domain-specific terminology, reasoning patterns, and factual knowledge across the breadth of Saudi professional life.

This sovereign data advantage is difficult to replicate. Private companies developing Arabic models must rely on publicly available web data, which skews heavily toward Modern Standard Arabic news content and informal social media text, leaving significant gaps in government, legal, medical, and technical Arabic. ALLaM’s access to internal government data — processed under SDAIA’s data governance framework — provides knowledge that no commercial alternative can match.

Platform Deployment

ALLaM’s deployment strategy reflects Saudi Arabia’s approach to technology partnerships: multi-platform availability that avoids vendor lock-in while leveraging the strengths of different cloud ecosystems.

IBM watsonx integration, announced in May 2024, made ALLaM available through IBM’s enterprise AI platform. Organizations can access the model through watsonx.ai studio, leveraging IBM’s governance capabilities for model training, fine-tuning, and deployment. This partnership provides enterprise-grade compliance and audit capabilities that are essential for regulated industries — banking, healthcare, and government — where model behavior must be traceable and explainable.

Microsoft Azure availability, launched in September 2024, brought ALLaM to the world’s second-largest cloud platform. ALLaM-2-7b-instruct was made accessible to Azure customers in Saudi Arabia and globally, with the model having been developed and trained on Azure infrastructure. The Azure integration connects ALLaM to Microsoft’s enterprise ecosystem — including Office 365, Dynamics, and Power Platform — enabling Arabic AI capabilities within the productivity tools that Saudi organizations already use daily.

Hugging Face distribution in early 2025 extended accessibility to the broader research and developer community. The ALLaM-7B-Instruct-preview made available through Hugging Face enables researchers to evaluate, fine-tune, and build upon the model without requiring cloud platform subscriptions.

HUMAIN Chat

HUMAIN Chat represents ALLaM’s consumer-facing deployment, providing direct access to the model’s capabilities through a web-based interface. The platform offers features specifically designed for Arabic users: real-time web search integration for up-to-date knowledge access, Arabic speech input supporting multiple dialectal varieties, seamless bilingual switching that allows users to alternate between Arabic and English within a single conversation, conversation sharing for collaborative use, and full compliance with Saudi Arabia’s Personal Data Protection Law.

The speech input capability is particularly significant. Arabic speech recognition remains technically challenging due to dialectal variation, and HUMAIN Chat’s support for multiple Arabic dialects in speech input positions it ahead of most competing Arabic AI interfaces that require typed text input. This voice capability makes ALLaM accessible to users who may not be comfortable typing in Arabic — a substantial population segment, particularly among older users and those who primarily use Arabic in spoken rather than written communication.

ALLaM Challenge

SDAIA launched the ALLaM Challenge as a mechanism for accelerating Arabic AI application development. The competition invited developers to create innovative applications using ALLaM across categories including Arabic poetry analysis and generation, sentence parsing and grammatical analysis, educational applications, and dialogue simulation. Winners were announced in October 2024 and awarded prizes worth up to SAR 1 million (approximately $267,000), with winning applications demonstrating ALLaM’s versatility across creative, analytical, and educational use cases.

The challenge serves dual purposes: generating practical applications that demonstrate ALLaM’s commercial viability, and building a developer community around the model that creates ecosystem momentum independent of SDAIA’s internal development efforts.

Strategic Positioning

ALLaM’s strategic significance extends beyond its technical capabilities. As the national Arabic language model of the world’s largest economy in the Arab world, it serves as the linguistic foundation for Saudi Arabia’s broader AI ambitions. The model’s integration into government services — potentially encompassing citizen interaction, document processing, legal analysis, and educational assessment — would create a pervasive Arabic AI presence that shapes how millions of people interact with the Saudi state.

The transition from SDAIA to HUMAIN management in 2025 reflects the model’s evolution from a research project to a commercial platform. HUMAIN’s mandate to become the world’s third-largest AI provider (behind the United States and China) positions ALLaM not as a Saudi-only model but as the foundation for a globally competitive Arabic AI platform. The $77 billion data center infrastructure investment, the $10 billion venture fund for AI startups, and the partnerships with NVIDIA, AMD, and AWS collectively create an ecosystem designed to make ALLaM the default Arabic language model for enterprise, government, and consumer applications across the Arabic-speaking world.

Cohere’s ranking of ALLaM 34B as the world’s most advanced Arabic LLM built in the Arab world on the MMLU benchmark validates this positioning, though the competitive landscape is rapidly evolving as Falcon-H1 Arabic, Jais 2, and other models continue to advance.

Benchmark Context and Native Arabic Evaluation

ALLaM’s benchmark positioning must be understood within the evolving evaluation landscape for Arabic AI. The Open Arabic LLM Leaderboard, launched in May 2024 by 2A2I, TII, and Hugging Face, now carries over 700 model submissions from more than 180 organizations. The leaderboard’s version 2 removed machine-translated tasks entirely, replacing them with four native Arabic benchmarks: ArabicMMLU (14,575 questions from Arabic educational exams), ALRAGE (retrieval-augmented generation evaluation), AraTrust (522 human-written questions across eight trustworthiness dimensions), and MadinahQA (Islamic and cultural knowledge evaluation).

This shift toward native benchmarks matters for ALLaM because the model’s training data — assembled from Saudi government entities, educational institutions, and subject matter experts — aligns directly with the knowledge domains these benchmarks test. ArabicMMLU’s questions span STEM, social sciences, humanities, and Arabic language at all school levels through university, covering territory where ALLaM’s institutional training data provides concrete advantage. AraTrust’s evaluation of truthfulness, ethics, privacy, and cultural sensitivity aligns with the compliance framework built into ALLaM 34B under HUMAIN’s development standards and Saudi Arabia’s Personal Data Protection Law requirements.

The BALSAM benchmark, with its 78 tasks and 52,000 samples featuring private test sets to prevent data contamination, provides additional evaluation rigor. SILMA AI’s Arabic Broad Benchmark, covering 22 categories with 470 human-validated questions from 64 Arabic datasets, adds further breadth. ALLaM’s performance across these diverse evaluations confirms that the model’s sovereign data advantage translates to measurable capability gains on tasks requiring genuine Arabic knowledge rather than surface-level pattern recognition.

Competitive Landscape

ALLaM operates in a three-way competition with Jais and Falcon Arabic that reflects broader UAE-Saudi technology rivalry. Jais 2, with 70 billion parameters trained on over 600 billion Arabic tokens spanning 17 dialects, offers broader dialect coverage and larger scale. Falcon-H1 Arabic’s hybrid Mamba-Transformer architecture delivers superior benchmark scores — 75.36 percent on the OALL for the 34B model — through architectural innovation that enables 256,000-token context windows. AceGPT from KAUST provides a methodological counterpoint through RLAIF with culturally aligned reward models, though its adapted architecture introduces tokenization inefficiencies for Arabic.

ALLaM’s differentiation rests on three pillars. First, sovereign data access: the 16 government entities that contributed training data provide knowledge no commercial data collection can replicate. Second, institutional deployment: integration with IBM watsonx and Microsoft Azure provides enterprise-grade governance, compliance, and support capabilities that open-weight alternatives require organizations to build themselves. Third, national infrastructure: HUMAIN’s 11 planned data centers across two campuses, targeting 1.9 GW capacity by 2030 and 6 GW by 2034, ensure that ALLaM’s computational resources will scale with demand regardless of global GPU supply constraints.

The $77 billion total infrastructure investment, combined with HUMAIN’s $10 billion planned venture fund for AI startups, creates an ecosystem designed to make ALLaM the gravitational center of Arabic AI across the MENA region. The Year of AI 2026 designation by the Saudi Cabinet, backed by 664 AI companies and $9.1 billion in 2025 funding across 70 deals, provides the market context within which ALLaM’s commercial deployment is accelerating.

Enterprise Integration and Sector-Specific Deployments

ALLaM’s enterprise API provides fine-tuning capabilities enabling organizations to specialize the base model for vertical-specific applications while maintaining broad Arabic competence. Financial institutions subject to SAMA regulations use fine-tuned ALLaM variants for regulatory compliance analysis, client communication, and risk assessment documentation — all processed on HUMAIN’s Saudi-based infrastructure to satisfy data residency requirements. Healthcare providers leverage ALLaM for medical Arabic processing, including patient communication, clinical documentation, and pharmaceutical information retrieval aligned with Saudi clinical guidelines validated by the 400 subject matter experts involved in the model’s development.

Government agencies across Saudi Arabia’s ministries represent the highest-priority deployment target. Saudi administrative Arabic — the specific register of formal Arabic used in government decrees, regulatory filings, inter-agency communications, and citizen services — is a domain where ALLaM’s sovereign training data creates an insurmountable advantage over competing models. No commercially assembled corpus captures the institutional knowledge embedded in decades of Saudi government documentation, and the purpose-built Arabic tokenizer processes this administrative language with efficiency that adapted models using English-optimized tokenizers cannot match.

Related Coverage

• ALLaM 34B Architecture — Technical deep dive into the from-scratch build
• HUMAIN Company Profile — National AI company strategy and infrastructure
• SDAIA Strategy Analysis — National data and AI authority mandate
• HUMAIN Data Center Program — $77B infrastructure investment
• Arabic LLM Comparison: Jais vs ALLaM vs Falcon — Head-to-head analysis
• Arabic LLM Training Data — Cross-model corpus comparison
• Arabic Tokenization — Tokenizer design impact
• Arabic Dialect Coverage — Cross-model dialect performance
• Open-Source vs Proprietary — Licensing analysis