The December 2025 release of Jais 2 marked a watershed moment for Arabic artificial intelligence. With 70 billion parameters trained on the richest Arabic-first dataset ever assembled — exceeding 600 billion Arabic tokens — the model represents four generations of accumulated learning by the G42-MBZUAI-Cerebras consortium.
The architectural redesign that underpins Jais 2 reflects hard-won insights from previous releases. The original Jais-13B demonstrated that dedicated Arabic training could produce models competitive with much larger multilingual systems. Jais-30B confirmed that parameter scaling delivered proportional quality improvements. The 2024 family release — 20 models spanning 590M to 70B parameters — provided the data needed to optimize training procedures across the full size spectrum. Jais 2 synthesizes these lessons into a single flagship model that pushes every dimension of Arabic AI capability forward.
Training Data Evolution
The training corpus for Jais 2 exceeds 600 billion Arabic tokens, representing a dramatic expansion from the 116 billion Arabic tokens used for the original Jais-13B. This growth reflects both the availability of more Arabic digital content and improved data collection and curation processes.
The corpus composition emphasizes diversity across multiple axes. Temporal diversity ensures representation of both historical and contemporary Arabic, preventing the model from developing a narrow temporal perspective. Geographic diversity covers all major Arabic-speaking regions, with deliberate oversampling of dialectal content that would otherwise be underrepresented relative to Modern Standard Arabic. Domain diversity spans news, academic, legal, medical, technical, literary, religious, and conversational content. Register diversity encompasses formal and informal communication, written and transcribed spoken language, and professional and personal contexts.
The Arabizi component — Arabic written in Latin characters — receives particular attention in Jais 2. This writing system, dominant in informal digital communication among younger Arabic speakers, was underrepresented in earlier training corpora. Jais 2 includes substantial Arabizi training data, enabling the model to process and generate this register naturally. For applications targeting younger demographics — social media analysis, customer service chatbots, educational platforms — Arabizi competence is essential.
Dialect Coverage Expansion
Jais 2 explicitly trains on 17 identified regional Arabic dialects, a significant expansion from earlier versions. Gulf Arabic varieties include UAE, Saudi, Kuwaiti, Bahraini, Qatari, and Omani speech patterns. Egyptian Arabic — the most widely understood dialect due to Egypt’s dominance in Arabic media — receives proportional representation. Levantine varieties cover Palestinian, Jordanian, Lebanese, and Syrian Arabic. Iraqi Arabic is treated as a distinct category. The Maghrebi group encompasses Moroccan, Algerian, Tunisian, and Libyan varieties. And Sudanese Arabic rounds out the dialect inventory.
Each dialect presents unique challenges for language modeling. Phonological differences that are transparent in speech create ambiguity in text. Lexical variation means the same concept may be expressed with entirely different words across dialects. Grammatical structures diverge significantly from MSA and from each other. Jais 2 addresses these challenges through dialect-aware training that preserves each variety’s linguistic characteristics.
Safety Framework
Jais 2 incorporates a comprehensive safety framework reflecting the cultural context of Arabic-speaking societies. The framework operates across multiple dimensions: factual accuracy to prevent misinformation, cultural sensitivity to avoid content violating social norms, religious awareness for handling Islamic references appropriately, political neutrality across the diverse landscapes of Arabic-speaking countries, and protection against generating harmful content.
The safety framework was developed in consultation with Arabic-speaking subject matter experts from multiple countries, ensuring that safety definitions reflect the diversity of Arabic cultural contexts rather than imposing a single perspective.
Performance Assessment
Jais 2 demonstrates measurable improvements across all evaluation dimensions. Reasoning tasks show the most dramatic improvement, reflecting increased parameter count and cleaner training data. Arabic fluency improvements are particularly pronounced in dialectal generation where earlier models mixed dialect features inappropriately. Bilingual performance remains strong — English benchmark scores are competitive with English-first models of similar size, confirming that Arabic-first training does not sacrifice English capability.
The Open Arabic LLM Leaderboard provides the most relevant evaluation framework for Jais 2. Launched in May 2024 by 2A2I, TII, and Hugging Face, the OALL’s version 2 benchmarks — ArabicMMLU, ALRAGE, AraTrust, and MadinahQA — evaluate exclusively on native Arabic content. ArabicMMLU’s 14,575 questions, sourced from educational exams across Arab countries covering STEM, social sciences, humanities, and Arabic language at all school levels through university, test knowledge depth that machine-translated benchmarks failed to capture. Jais 2’s strong ArabicMMLU performance reflects the breadth of its 600B+ Arabic training corpus across academic, professional, and cultural domains.
AraTrust evaluation is particularly revealing for Jais 2. The benchmark’s 522 human-written questions assess eight trustworthiness dimensions: truthfulness, ethics, privacy, illegal activities, mental health, physical health, unfairness, and offensive language. Earlier evaluations showed that smaller Arabic models — including AceGPT 7B and Jais 13B — scored below 60 percent on AraTrust, revealing that accuracy and trustworthiness are distinct capabilities. Jais 2’s comprehensive safety framework, developed with Arabic-speaking experts from multiple countries, addresses these dimensions with the maturity that four generations of development provide.
BALSAM, comprising 78 tasks and 52,000 samples with private test sets preventing data contamination, provides additional evaluation rigor that protects against benchmark overfitting. SILMA AI’s Arabic Broad Benchmark, with 470 human-validated questions from 64 Arabic datasets across 22 categories, adds breadth. The evaluation ecosystem now includes over 40 distinct Arabic benchmarks across LLM performance, multimodality, embedding, retrieval, RAG generation, speech, and OCR — and Jais 2 is evaluated against this full spectrum.
Architecture Deep Dive
Jais 2’s decoder-only transformer architecture incorporates several refinements over the original Jais-13B design. The tokenizer was rebuilt to maximize Arabic efficiency, treating common morphological patterns — prefixed conjunctions, prepositional clitics, pronominal suffixes, and definite articles — as single tokens rather than fragmenting them into character sequences. This tokenization efficiency means equivalent Arabic text requires fewer tokens in Jais 2 than in adapted models like AceGPT, which inherits Llama 2’s English-optimized tokenizer and processes Arabic at the individual letter level.
The attention mechanism was optimized for Arabic’s syntactic characteristics. Arabic’s VSO (verb-subject-object) word order places critical grammatical information at sentence beginnings, requiring the model to maintain this context across the entire sequence. Complex morphological agreement patterns — where verbs agree with subjects in person, number, and gender across potentially long distances — demand attention patterns distinct from those optimal for English’s SVO structure. Jais 2’s attention architecture was trained from scratch on Arabic syntax, avoiding the English-biased attention patterns inherited by adapted models.
The 70-billion parameter scale was selected based on scaling law analysis from the 2024 family release. The consortium’s experience training 20 models spanning 590M to 70B parameters provided empirical data on the relationship between parameter count, training data volume, and downstream task performance for Arabic specifically. This data indicated that the 70B scale, combined with 600B+ Arabic training tokens, crosses capability thresholds on reasoning and dialectal generation tasks that smaller models do not reach regardless of training data quality.
Competitive Benchmarking
Jais 2 competes directly with Falcon-H1 Arabic’s 34B model, which achieves 75.36 percent on the OALL through its hybrid Mamba-Transformer architecture. The architectural comparison reveals a fundamental design tradeoff: Jais 2’s pure transformer architecture provides proven reasoning capabilities and established fine-tuning tooling, while Falcon-H1’s hybrid design achieves competitive performance at lower parameter counts through the Mamba state-space model’s linear sequence processing efficiency.
ALLaM 34B from HUMAIN presents a different competitive dimension. Built from scratch with Arabic-optimized tokenization and trained on sovereign data from 16 Saudi government entities, ALLaM 34B was ranked by Cohere as the world’s most advanced Arabic LLM built in the Arab world on the MMLU benchmark. ALLaM’s deployment through IBM watsonx and Microsoft Azure provides enterprise governance capabilities that Jais’s open-weight distribution does not include natively, though the open-weight model can be deployed with third-party governance solutions.
The three-way competition benefits the Arabic AI ecosystem by ensuring no single model dominates, driving rapid innovation across architecture design, training data curation, and benchmark evaluation methodology simultaneously.
Training Infrastructure
The computational foundation for Jais 2 is the Condor Galaxy 1 supercomputer, built jointly by G42 and Cerebras Systems. The system’s multi-exaFLOP performance, based on Cerebras CS-2 wafer-scale engines integrating 850,000 AI-optimized compute cores onto single chips, eliminates the communication bottlenecks that limit conventional GPU cluster efficiency. For Jais 2’s training campaign — 70 billion parameters on 600+ billion Arabic tokens — the wafer-scale architecture provided sustained utilization rates exceeding those achievable on distributed GPU clusters, reducing actual training costs below what headline FLOPS comparisons would suggest.
The sovereign computing infrastructure ensures that Jais’s development capacity is independent of foreign cloud providers or GPU supply constraints. As the UAE positions itself as a global center for Arabic AI development, this computational sovereignty — training frontier models without external dependencies — complements the data sovereignty represented by the Arabic-first training corpus.
Deployment Ecosystem and Developer Access
Jais 2 is distributed through multiple access channels designed to maximize adoption across developer segments. Hugging Face hosts the open-weight models — Jais-2-8B-Chat and Jais-2-70B-Chat — with full model weights, tokenizers, and configuration files available for download and local deployment. The JaisChat.ai web interface provides direct interaction without infrastructure requirements, enabling non-technical users to evaluate the model’s capabilities before committing to API integration or self-hosted deployment.
The open-weight licensing model differentiates Jais from ALLaM’s platform-dependent distribution through IBM watsonx and Microsoft Azure. Organizations deploying Jais 2 can run inference on their own infrastructure, fine-tune the model on proprietary datasets, integrate with any orchestration framework, and modify the model architecture without licensing restrictions. This flexibility is particularly valuable for MENA startups operating within the region’s $858 million AI venture capital ecosystem, where computational budgets are finite and vendor lock-in carries strategic risk.
G42’s partnership with Microsoft provides Azure deployment options for organizations that prefer managed cloud infrastructure. The $2.3 billion Microsoft investment in G42 ensures long-term platform integration, though the open-weight license means that Azure is an option rather than a requirement. This dual-access model — open weights for maximum flexibility, Azure integration for managed deployment — addresses both the startup and enterprise segments of the MENA AI market.
Implications for Arabic AI Research
Jais 2’s open-weight release has catalyzed Arabic AI research by providing a high-quality baseline model that academic and commercial researchers can study, evaluate, and extend. The model’s architecture, training methodology, and performance characteristics are documented in detail, enabling the kind of reproducible research that proprietary models preclude. Research groups at MBZUAI, KAUST, NYU Abu Dhabi’s CAMeL Lab, and universities across the MENA region use Jais 2 as a foundation for dialect-specific fine-tuning, domain adaptation, and novel training methodology experiments.
The four-generation development lineage from Jais-13B through Jais 2 provides a longitudinal case study in Arabic LLM scaling that no other model family offers. Researchers can trace how architectural decisions, training data composition, and safety frameworks evolved across generations, extracting principles applicable to Arabic AI development broadly. This research transparency accelerates the entire field, contributing to the ecosystem growth that Saudi Arabia’s SDAIA and the UAE’s national AI strategies both target as strategic priorities.
Serving Infrastructure and Inference Optimization
Jais 2 is both trained and served on Cerebras systems, with G42 reporting that their ML techniques uniquely enabled by Cerebras hardware achieved state-of-the-art quality using only a fraction of the compute used to train similar-sized models in the past. The inference deployment architecture leverages the same wafer-scale engine advantages that benefit training — high memory bandwidth, on-chip communication, and sustained throughput at long sequence lengths — to deliver competitive per-query latency for Arabic text generation. For enterprise deployments processing millions of daily queries across customer service, document analysis, and content generation workloads, this inference efficiency determines the economic viability of Jais 2 relative to smaller models that sacrifice quality for speed.
The dual deployment pathway through open weights and managed cloud reflects market segmentation. Startups and research organizations download weights from Hugging Face and deploy on their own GPU infrastructure, maintaining maximum control over model behavior and data flows. Enterprise customers preferring managed deployment access Jais 2 through Azure, leveraging Microsoft’s $2.3 billion investment in G42 for seamless platform integration. This flexibility positions Jais 2 for adoption across the full MENA market spectrum — from individual developers building proof-of-concept Arabic AI applications to government agencies deploying sovereign AI infrastructure at scale.
Related Coverage
- Jais — Complete Model Profile — Full Jais family analysis
- Condor Galaxy Supercomputer — Training infrastructure
- Arabic Dialect Coverage — Cross-model dialect comparison
- ALLaM 34B Architecture — HUMAIN’s from-scratch competitor
- Falcon-H1 Architecture — Hybrid Mamba-Transformer comparison
- OALL Benchmark Analysis — Leaderboard evaluation methodology
- G42 Company Profile — Corporate strategy and partnership ecosystem
- Arabic LLM Training Data — Cross-model corpus comparison