Jais — The World's Leading Arabic Open-Weight Large Language Model

Jais stands as the most significant contribution to Arabic-language artificial intelligence in the history of the field. Named after Jebel Jais, the highest peak in the United Arab Emirates, this open-weight large language model was developed through a trilateral partnership between G42’s Inception unit, the Mohamed bin Zayed University of Artificial Intelligence, and Cerebras Systems — a collaboration that united Emirati sovereign ambition, academic research excellence, and American supercomputing hardware into a single program that has fundamentally altered the competitive landscape of Arabic AI.

The model’s genesis in August 2023 addressed a structural gap that had persisted for the entire generative AI era. While English-language models had advanced rapidly from GPT-2 through GPT-4, Arabic speakers — numbering over 400 million across 22 countries — were served primarily by multilingual models that allocated fewer than two percent of training tokens to Arabic. The resulting systems could produce grammatically acceptable Modern Standard Arabic but failed catastrophically on regional dialects, cultural references, code-switching patterns, and the subtle pragmatic conventions that govern Arabic communication in professional, educational, and social contexts.

Development Timeline

The Jais development program has proceeded through four distinct phases, each representing a significant capability expansion.

Jais-13B (August 2023) launched as the inaugural release, establishing the foundational architecture. This 13-billion parameter model was trained on a purpose-built dataset comprising 116 billion Arabic tokens designed to capture the complexity, nuance, and richness of the language, supplemented by 279 billion English tokens to enable cross-lingual transfer learning. The Arabic dataset was constructed through a meticulous curation process that assembled content from news sources, literary works, educational materials, government documents, social media, and web crawls — filtered through quality classifiers to remove machine-translated content and ensure authentic Arabic representation.

Jais-30B (November 2023) doubled the model’s capacity within three months of the initial launch. The 30-billion parameter version demonstrated that scaling delivered proportional improvements in Arabic language understanding, particularly in tasks requiring multi-step reasoning, extended text generation, and domain-specific knowledge retrieval. The rapid release cadence signaled G42’s commitment to continuous development rather than the long gaps between versions that characterized some competing projects.

Jais Family Release (H2 2024) represented the most ambitious single model release in MENA history. The consortium published 20 open-source Arabic-centric models spanning from 590 million to 70 billion parameters, trained on up to 1.6 trillion tokens of Arabic, English, and code data. This range addressed the full deployment spectrum — from edge devices and mobile applications (sub-1B models) through enterprise workloads (7B-13B) to research and premium applications (30B-70B). The release strategy explicitly targeted the developer ecosystem, providing multiple entry points for teams with varying computational resources.

Jais 2 (December 2025) arrived as the definitive Arabic-first foundation model. Built from the ground up with 70 billion parameters and trained on the richest Arabic-first dataset assembled to date — exceeding 600 billion Arabic tokens — Jais 2 incorporated a redesigned architecture, cleaner training data pipelines, and enhanced safety frameworks. The model demonstrates stronger reasoning capabilities, greater fluency across Modern Standard Arabic and 17 regional dialects, and competitive English performance despite its Arabic-first design philosophy.

Architecture and Training

Jais is a generative pre-trained transformer model following the decoder-only architecture that has become standard for large language models. The training infrastructure centers on Condor Galaxy 1, the multi-exaFLOP AI supercomputer built jointly by G42 and Cerebras. This system, based on Cerebras CS-2 wafer-scale engines, provides the computational throughput necessary for training models at the 70-billion parameter scale within commercially viable timeframes.

The training data composition for Jais 2 reflects lessons learned from four previous releases. The Arabic corpus exceeds 600 billion tokens drawn from multiple sources: Modern Standard Arabic from news organizations, academic publications, and government documents; dialectal Arabic from social media platforms, online forums, and transcribed broadcast content; classical Arabic from literary and religious texts; and Arabizi — the widespread practice of writing Arabic using Latin characters — from messaging platforms and informal digital communication. The dataset explicitly includes 17 identified regional dialects, ensuring that the model can process and generate text in Gulf Arabic, Egyptian Arabic, Levantine Arabic, Maghrebi Arabic, and other major dialect groups.

The English component serves a dual purpose. Beyond enabling bilingual operation, English training data provides knowledge transfer benefits: concepts and reasoning patterns learned from the larger English corpus transfer to Arabic tasks through the model’s shared parameter space. Research published by the Jais team demonstrated that this cross-lingual transfer operates bidirectionally — the model’s English capability benefits from Arabic training data patterns and vice versa.

Dialect Coverage and Cultural Competence

The breadth of dialect coverage distinguishes Jais from competing Arabic models. While most Arabic LLMs achieve strong performance on Modern Standard Arabic — the formal register used in news, academic writing, and official communication — performance typically degrades significantly on dialectal Arabic. Jais 2 addresses this through explicit training on 17 regional dialect varieties, enabling the model to navigate the linguistic diversity that characterizes real-world Arabic communication.

The dialect coverage encompasses Gulf Arabic (spanning UAE, Saudi, Kuwaiti, Bahraini, Qatari, and Omani varieties), Egyptian Arabic (the most widely understood dialect due to Egypt’s media influence), Levantine Arabic (Palestinian, Jordanian, Lebanese, and Syrian varieties), Iraqi Arabic, Maghrebi Arabic (Moroccan, Algerian, Tunisian, and Libyan varieties), and Sudanese Arabic. Additionally, the model processes Arabizi with contextual accuracy, recognizing that this romanized Arabic writing system is the dominant mode of informal digital communication for younger Arabic speakers.

Cultural competence extends beyond linguistic accuracy. Jais 2 incorporates knowledge of Arabic poetry meters and forms, cultural references specific to Arab societies, social communication norms, religious terminology and concepts, and the pragmatic conventions that govern Arabic discourse. The comprehensive safety framework ensures that generated content aligns with the cultural expectations of Arabic-speaking users while maintaining factual accuracy and avoiding harmful outputs.

Benchmark Performance

On native Arabic benchmarks — the evaluations that matter most for real-world Arabic deployment — Jais consistently demonstrates strong performance. The model outperforms existing Arabic models by a sizable margin across tasks including reading comprehension, question answering, sentiment analysis, text classification, and natural language inference. Notably, Jais achieves competitive results on English benchmarks despite the deliberate Arabic-first training priority, confirming the bidirectional transfer learning thesis.

On the Open Arabic LLM Leaderboard, Jais models rank among the top performers in their respective size categories. The OALL, launched in May 2024 by 2A2I, TII, and Hugging Face, has received over 700 model submissions from more than 180 organizations worldwide. The leaderboard’s version 2 benchmarks — ArabicMMLU, ALRAGE, AraTrust, and MadinahQA — are all native Arabic evaluations, removing the machine-translated tasks that inflated scores for models trained on translated content. ArabicMMLU alone comprises 14,575 native Arabic multiple-choice questions sourced from educational exams across Arab countries, covering all school levels through university across STEM, social sciences, humanities, and Arabic language domains.

The AraTrust benchmark adds a trustworthiness dimension absent from pure accuracy evaluations. With 522 human-written multiple-choice questions evaluating eight dimensions — truthfulness, ethics, privacy, illegal activities, mental health, physical health, unfairness, and offensive language — AraTrust revealed that some high-scoring models on accuracy benchmarks perform poorly on safety and alignment. Jais 2’s comprehensive safety framework, developed in consultation with Arabic-speaking experts from multiple countries, addresses this dimension explicitly.

SILMA AI’s Arabic Broad Benchmark provides additional validation through 470 human-validated questions drawn from 64 Arabic datasets across 22 categories, with evaluation using over 20 manual rules combined with LLM-as-Judge variations per skill. BALSAM, with its 78 tasks and 52,000 samples including private test sets that prevent data contamination, offers yet another evaluation angle. Across this expanding battery of native Arabic evaluations, Jais maintains consistently strong positioning.

Availability and Licensing

Jais models are distributed under open-weight licenses through Hugging Face, with Jais-2-8B-Chat and Jais-2-70B-Chat available for download. The open-weight approach — making model weights available while retaining some usage restrictions — balances accessibility with responsible deployment. Developers can fine-tune and deploy Jais models for commercial applications, while the license terms discourage harmful use cases. This approach contrasts with ALLaM’s multi-platform availability through IBM watsonx and Microsoft Azure, which introduces enterprise licensing layers, and with Falcon’s Apache 2.0-based TII Falcon License, which is the most permissive among major Arabic LLMs, allowing virtually any commercial use without attribution requirements.

JaisChat.ai provides a web-based chat interface for direct interaction with the latest Jais models, serving as both a demonstration platform and a consumer-facing product. The chat interface supports Arabic and English input with seamless language switching, enabling bilingual users to interact naturally without explicit language selection. This consumer-facing deployment parallels HUMAIN Chat, which provides ALLaM access with additional features including real-time web search, multi-dialect speech input, and Saudi PDPL compliance.

Strategic Significance

Jais is not merely a technical achievement; it is a statement of sovereign AI capability. The UAE’s investment in Jais — through G42’s development resources, MBZUAI’s research capacity, and the Condor Galaxy computing infrastructure — positions the Emirates as a global center for Arabic AI development. The open-weight licensing ensures that this capability is available to the broader Arabic-speaking world, creating a network effect that strengthens the UAE’s position as the ecosystem hub.

The model’s significance is amplified by the broader G42 strategy. Microsoft’s $2.3 billion investment in G42 in 2024, combined with Jais’s deployment on Azure and integration with enterprise platforms, creates a commercial pathway that transforms research excellence into economic value. For the UAE, Jais demonstrates that sovereign AI development can be simultaneously world-class, open-source, and commercially viable — a template that other nations are studying closely.

Competitive Positioning

The Arabic LLM landscape as of early 2026 features three primary competitors, each with distinct advantages. Jais competes directly with ALLaM 34B, developed by HUMAIN (formerly SDAIA’s NCAI), which was trained on 500 billion Arabic tokens assembled from 16 Saudi government entities and 300 Arabic books reviewed by 400 subject matter experts. ALLaM’s sovereign data access — internal government documents, regulatory texts, and administrative records — provides knowledge that no commercially assembled corpus can replicate. However, ALLaM’s initial dependence on Meta’s Llama 2 architecture constrained its Arabic tokenization efficiency, a limitation only addressed with the from-scratch ALLaM 34B release in 2025.

Falcon-H1 Arabic from TII presents a different competitive challenge. The hybrid Mamba-Transformer architecture enables 256,000-token context windows that would be computationally prohibitive for Jais’s pure transformer design. Falcon-H1’s 34B model achieves 75.36 percent on the OALL, exceeding some models with more than 70 billion parameters. The architectural innovation — combining Mamba state-space model efficiency with transformer reasoning capability — positions Falcon as the performance leader on benchmark evaluations.

AceGPT from KAUST and CUHKSZ occupies a methodological niche, pioneering RLAIF with culturally aligned reward models for Arabic. While AceGPT demonstrated that cultural alignment could be engineered through reinforcement learning, its adaptation from Llama 2 introduces tokenization inefficiencies — processing Arabic text at the individual letter level rather than subword granularity — that increase inference costs relative to Arabic-native architectures.

Jais’s competitive advantage lies in the combination of scale (70B parameters), data breadth (600B+ Arabic tokens spanning 17 dialects), open-weight accessibility, and the sovereign computing infrastructure that ensures continued development capacity. The Condor Galaxy 1 supercomputer provides training throughput that makes iterative model improvement economically feasible, while the open-weight licensing strategy builds ecosystem momentum that proprietary alternatives cannot match.

Developer Ecosystem and Community Adoption

The open-weight distribution strategy has generated substantial developer ecosystem momentum. Hugging Face downloads for Jais models number in the millions, with developers across the MENA region building applications that span customer service automation, educational content generation, legal document analysis, healthcare information systems, and creative writing tools. The developer community contributes fine-tuning scripts, evaluation code, integration examples for LangChain, LangGraph, and CrewAI orchestration frameworks, and dialect-specific model adaptations that extend Jais’s capabilities beyond what the base model provides.

The MENA AI startup ecosystem increasingly builds on Jais as a foundation. With $858 million in AI-focused venture capital deployed in the region during 2025 — representing 22 percent of total VC funding — startups selecting Arabic LLM foundations evaluate Jais’s open-weight accessibility, Falcon’s Apache 2.0 licensing, and ALLaM’s enterprise platform integration. Jais’s combination of open weights, large parameter count, and broad dialect coverage makes it the preferred choice for startups requiring maximum deployment flexibility without licensing constraints. The $10 billion HUMAIN venture fund and $1 billion GAIA Accelerator create additional ecosystem capital, though these funds naturally favor ALLaM-based applications within the Saudi market.

Related Coverage

• Jais 2 Deep Analysis — Technical deep dive into the December 2025 release
• Condor Galaxy Supercomputer — Training infrastructure analysis
• G42 Company Profile — Corporate strategy and AI portfolio
• MBZUAI Research Profile — Academic research contributions
• Arabic LLM Training Data — Cross-model training corpus comparison
• OALL Benchmark Analysis — Leaderboard performance context
• ALLaM — Saudi Arabia’s National Model — Primary competitor analysis
• Falcon Arabic — TII’s Hybrid Architecture — Benchmark leader comparison
• AceGPT — Cultural Alignment — Methodological alternative
• Arabic vs English LLM Performance — Cross-language capability gap