Top Arabic-Language LLMs Available in Saudi Arabia

The Arabic Language Challenge: Why It’s Genuinely Hard

Any sophisticated analysis of Arabic LLMs must begin with an honest assessment of why Arabic NLP is structurally more difficult than English NLP — and why this difficulty creates genuine moats for teams that solve it well.

Arabic is a morphologically rich language with root-based derivation, meaning a single three-letter root can generate dozens of derived word forms through patterns of vowels, prefixes, and suffixes. The word “يستخدمونه” (they are using it) encodes subject, tense, mood, object pronoun, and verb stem in a single token — creating a tokenization challenge that English-optimized tokenizers handle poorly. Modern Standard Arabic (MSA) and the numerous spoken dialects (Gulf, Egyptian, Levantine, Moroccan, and others) are mutually partially intelligible but linguistically distinct — a model trained predominantly on MSA may perform poorly on Gulf dialect inputs, which dominate Saudi social media and everyday digital communication.

Right-to-left text processing, bidirectional text in documents mixing Arabic and Latin characters, the absence of vowel diacritics in most written Arabic (requiring contextual disambiguation), and the significant difference between formal written and colloquial spoken Arabic collectively make Arabic NLP genuinely hard. The benchmark performance gap between leading Arabic LLMs and English LLMs, on equivalent tasks, remains meaningful even in 2025.

This difficulty creates the market opportunity. The 400 million+ Arabic-speaker population represents a genuine underserved AI market — one where the quality gap between Arabic-first models and English models with Arabic capability is significant enough to create sustained demand for specialized Arabic LLMs.

Allam: Saudi Arabia’s Sovereign Arabic Model

SDAIA’s Allam model is the most significant Arabic LLM developed within Saudi Arabia and one of the most significant Arabic-first models globally. At 34 billion parameters, trained on 8 petabytes of Arabic training data (one of the largest Arabic training datasets assembled), and deployed on 5,000 NVIDIA Blackwell GPUs at SDAIA’s Hexagon Data Center (480 MW), Allam represents a genuine sovereign AI capability.

The 8 PB training dataset is the critical differentiator. Arabic internet text has historically been underrepresented in large language model training corpora — GPT-3’s training data was approximately 7% Arabic by token count (generous estimates), and most Arabic tokens in English-first training sets were crawled from a narrow range of websites with limited dialect diversity. SDAIA assembled its 8 PB corpus through an extensive Arabic data collection program: crawling Arabic-language websites across all 22 Arab countries, digitizing historical Arabic texts and literature, incorporating Saudi government documents, and collecting Arabic social media data with dialect diversity.

The implications of scale Arabic training data are not just benchmark performance — they are contextual grounding. Allam understands references to Saudi institutions (the Absher platform, the Tadawul, the Ministry of Justice’s Najiz platform), Saudi cultural practices, Islamic legal terminology in Saudi context, and Saudi social norms in ways that emerge from data rather than rules. This contextual grounding is what makes Allam suitable as a foundation for Saudi-specific agentic AI applications.

Allam’s commercial availability — through SDAIA’s API and Humain’s infrastructure — means that Saudi enterprises and government entities can build Arabic-first AI applications on a domestically-produced model with no data sovereignty concerns. This is strategically significant for sensitive applications: legal AI, healthcare AI, government service automation, financial AI.

Competition: International Models with Arabic Capability

The global leading LLMs all have Arabic language capability, and for many applications their Arabic performance is already strong:

GPT-4o (OpenAI/Microsoft) performs well on Modern Standard Arabic tasks and has reasonable Gulf dialect comprehension. Its training data, while English-dominant, includes substantial Arabic content. For enterprise Arabic applications that do not require data sovereignty guarantees, GPT-4o through Azure OpenAI Service is a viable option — and Microsoft’s Humain partnership creates a specific pathway for GPT-4o deployment in Saudi enterprise contexts.

Claude Sonnet (Anthropic) has demonstrated strong Arabic language performance in independent benchmarks, with particular strengths in Arabic instruction following and structured output generation. Claude’s Constitutional AI approach may also be valuable in Saudi context — content filtering calibrated for Arabic cultural norms is different from English cultural norms.

Gemini Ultra/Pro (Google) brings substantial Arabic capability through Google’s multilingual training program. Google Translate’s Arabic data — covering formal and colloquial Arabic across all major dialects — provides a training data advantage for Arabic comprehension even if the primary model is not Arabic-first.

Meta Llama Arabic fine-tunes: The open-source nature of the Llama model family has spawned multiple Arabic fine-tunes by academic and commercial teams. AceGPT (from the National University of Singapore, Arabic fine-tune), AraLLaMA (from various Arabic NLP research groups), and Qwen2-Arabic (Alibaba, with strong Arabic performance) are all meaningful competitors in the open-source Arabic LLM space.

Mistral Arabic: Mistral’s efficient model architecture has been fine-tuned for Arabic by both Mistral itself and third-party teams. Mistral’s European origin (and associated GDPR-positive positioning) makes it attractive for Saudi healthcare and financial applications that require non-US data residency for compliance reasons.

Jais: The UAE Competitor

The most direct competition for Allam comes from the UAE, not from the US. Jais, developed by G42 and Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi, is a 70B-parameter Arabic-first model trained on extensive Arabic and English data. At 70B parameters vs. Allam’s 34B, Jais has a parameter size advantage — though parameter count is an increasingly imperfect proxy for actual model capability.

Jais’s competitive positioning relative to Allam reflects the broader Saudi-UAE AI competition. The UAE has moved faster in several dimensions: G42’s international partnerships (Microsoft investment, OpenAI engagement), MBZUAI’s research output, and the Jais model’s open availability (the model weights are publicly released, allowing external fine-tuning and evaluation) have given Jais significant international profile.

For Saudi-specific applications, Allam has the Saudi contextual grounding advantage. For pan-Arabic applications serving audiences across all Arab countries, the competition between Allam and Jais is genuine. The Arabic LLM market is not winner-take-all — regional and use-case specialization means both models will likely find deployment niches.

The Saudi-UAE dimension creates an interesting dynamic: these are close strategic partners (both GCC members, both Vision-2030-adjacent economies), but also direct competitors for AI infrastructure primacy in the Middle East. The Allam-Jais competition is one expression of this broader competition.

Arabic Language Model Benchmarks: What Matters

Standard English-language LLM benchmarks (MMLU, HellaSwag, HumanEval) are poor evaluators of Arabic LLM performance. Several Arabic-specific benchmarks are now established:

ArabicMMLU is the Arabic adaptation of the Massive Multitask Language Understanding benchmark, covering 40 subjects translated and adapted for Arabic context. Allam scores above 70% on ArabicMMLU — competitive with GPT-4o-mini and above most open-source Arabic models, though below GPT-4o and Claude Sonnet on this benchmark.

ArabicBench covers Arabic reading comprehension, question answering, and summarization across MSA and dialect variants. The dialect variants are the critical differentiator — models that perform well on MSA but poorly on dialect text are not practically useful for Saudi consumer applications.

AraGPT2 perplexity benchmarks measure Arabic fluency — how well a model predicts Arabic text. Perplexity is a fundamental measure of Arabic language mastery that correlates with downstream task performance.

Cultural alignment benchmarks are the newest and arguably most important dimension: does the model’s responses reflect culturally appropriate perspectives on Islamic finance, gender norms, family structure, and Saudi cultural practices? Allam’s training data gives it inherent advantages here that parameter-scale advantages of foreign models cannot easily compensate.

The Business Case for Arabic-First AI

The 400+ million Arabic speaker market represents substantial and underserved commercial opportunity. Arabic digital content consumption is among the highest per-capita globally — Saudi Arabia has one of the world’s highest social media engagement rates, with Twitter/X and Instagram penetration among the top globally. Yet Arabic-language AI products — customer service bots, content generation tools, AI search — have historically been significantly inferior to English equivalents.

The commercial opportunity is largest in:

Arabic content generation: Marketing copy, social media content, news article drafting, and legal document generation in Arabic are high-volume enterprise use cases with clear ROI. Companies that can deliver high-quality Arabic content at scale have significant competitive advantage in Gulf markets.

Arabic customer service AI: The Saudi banking, telecoms, and retail sectors collectively serve tens of millions of Arabic-speaking customers. A customer service AI that genuinely understands Arabic — including dialect, slang, and contextually implicit meaning — vs. one that processes Arabic as “noisy English” delivers meaningfully different customer experience and deflection rates.

Arabic education technology: The Saudi education system is Arabic-medium at primary and secondary levels, and Arabic language AI tutors, automated essay evaluation, and Arabic reading assistance tools address an enormous and underserved market.

Arabic legal and regulatory AI: Saudi Arabian law is documented primarily in Arabic, including court decisions, Royal Decrees, and regulatory circulars. AI tools for legal research, compliance monitoring, and contract analysis in Arabic are high-value, high-demand applications that require Arabic-first models.

Sovereign vs. Commercial Arabic AI Strategy

The tension in Saudi Arabic AI strategy is between sovereign control and quality access. Allam, as a domestically-produced model, ensures that Saudi AI applications are not dependent on foreign model providers — a genuine sovereignty concern given that OpenAI, Anthropic, and Google models require data routing through US infrastructure with potential US government access.

But commercial quality considerations push toward international models for many applications. GPT-4o Arabic is currently better than Allam on many standardized benchmarks, and enterprise customers optimizing for capability rather than sovereignty will use it unless compelled otherwise.

The resolution being pursued in Saudi Arabia is layered: sovereign infrastructure (Humain compute, Hexagon data centers) hosting international model weights through licensing agreements, combined with Allam serving as the sovereign fallback and sovereign-required application layer. This “AI sovereignty through infrastructure control, not model exclusivity” approach allows Saudi enterprises to access GPT-4o quality through Saudi infrastructure — satisfying both capability and sovereignty requirements.

The Multimodal Arabic AI Frontier

The Arabic LLM competition is moving beyond text-only models. The next competitive frontier is multimodal Arabic AI — models that can process Arabic text alongside Arabic-script documents (legal contracts, medical records, engineering drawings with Arabic annotations), Arabic audio (dialect speech recognition), and Arabic video content. Each of these modalities has both a general AI challenge and a specific Arabic challenge.

Arabic speech recognition has historically underperformed English ASR because Arabic dialect diversity creates data collection challenges — training a model that handles Saudi Gulf dialect, Egyptian dialect, and Moroccan Darija simultaneously requires dialect-diverse audio training data that does not exist at scale. Saudi Telecom Company (stc) and the major Gulf telecoms have enormous Arabic audio datasets from call center recordings — datasets that are technically available for Arabic ASR fine-tuning but that raise privacy and PDPL compliance questions. SDAIA’s data governance authority over national datasets creates a potential pathway for aggregating dialect audio data for Allam’s multimodal extension, if the regulatory framework supports it.

Arabic document understanding (reading contracts, forms, and structured documents in Arabic script) is a near-term commercial priority for legal, healthcare, and financial applications. Arabic documents often mix formal Arabic, numbers, and Latin characters (transliterated names, technical terms) in ways that require robust handling of bidirectional text and mixed-script contexts. Models that handle this well have immediate commercial applications in Saudi legal processing, healthcare record digitization, and financial document automation.

The Global Arabic LLM Market: Beyond Saudi Arabia

Saudi Arabia’s investment in Arabic LLMs is strategically valuable not just for the domestic market but for the broader 400 million+ Arabic speaker global market. Egyptian Arabic speakers represent 100 million people. Iraqi, Syrian, Jordanian, Moroccan, and other Arab country populations collectively dwarf Saudi Arabia’s 35 million population. Allam and its derivatives are positioned to serve this broader market — creating export potential for Saudi AI that parallels Saudi Arabia’s energy export model.

The Gulf sovereign wealth funds (PIF, ADQ, Mubadala) are collectively building an Arabic AI ecosystem — with some competition (Allam vs. Jais) but also potential for regional coordination. A pan-Gulf Arabic LLM consortium, potentially building on both Allam and Jais architectures with combined training data from Saudi, UAE, Egyptian, and other Arab country sources, would produce a model competitive with English-language frontier models on Arabic tasks in ways that individual sovereign efforts cannot. Whether the Saudi-UAE AI competition allows this kind of collaboration is an open question that will likely be resolved by political dynamics above the technical level.

What Matters Beyond Parameter Count

For practitioners evaluating Arabic LLMs, the key dimensions beyond parameter count are: Arabic tokenization efficiency (tokens per Arabic character ratio, which affects inference cost and context window utilization), dialect coverage (can the model handle Gulf, Egyptian, and Levantine dialect inputs?), Islamic knowledge and Sharia comprehension (critical for financial and legal applications), Saudi cultural alignment (does the model’s default behavior match Saudi social norms?), and fine-tuning efficiency (can domain-specific fine-tunes be created cost-effectively?). Allam’s design prioritized all of these dimensions, and this holistic optimization for Arabic-specific requirements is more important than raw parameter count for the Saudi deployment context.

The investment implication for companies building on Arabic LLMs is to evaluate models not just on standard benchmarks but on the specific task distributions of the target application. A model that scores 72% on ArabicMMLU but performs at 90%+ on Gulf dialect customer service inputs may be significantly more valuable for a Saudi call center AI application than a model scoring 78% on ArabicMMLU but with lower dialect robustness. Task-specific evaluation on Saudi-specific test sets — which any serious Arabic LLM vendor should be able to provide — is the correct assessment methodology.