News
Feb 19, 2026
Tech Updates
Open-Source
Artificial Intelligence
Middle East & Africa
Jais 2 Delivers Frontier Arabic AI at 2,000 Tokens Per Second
NewDecoded
4 min read
Image by MBZUAI
State-of-the-Art Performance for 400 Million Speakers
G42's Inception, the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), and Cerebras Systems released Jais 2 on December 9, 2025, marking a breakthrough for Arabic language AI. The model family includes 8B and 70B parameter variants trained on 2.6 trillion tokens, with the larger model setting new records on AraGen, a leading Arabic benchmark. Jais 2 excels not only in general tasks like translation and summarization but also in culturally specific domains including Arabic poetry, religious reasoning, cuisine, and dream interpretation.
Unprecedented Speed Through Wafer-Scale Architecture
The Jais 2 70B chat application achieves production speeds of 2,000 tokens per second, making it approximately 20 times faster than frontier models like GPT-4 and Claude on real workloads. This performance stems from Cerebras' CS-3 inference architecture, which loads all model weights directly into on-wafer SRAM with petabyte-per-second bandwidth. Combined with custom speculative decoding, the system enables instant document summarization, real-time code iteration, and low-latency voice agents.
Efficient Training on Unified Memory Clusters
Unlike GPU-based systems requiring hundreds of interconnected chips, Jais 2 trained end-to-end on Condor Galaxy clusters containing 64 Cerebras CS-2 systems connected to unified MemoryX fabric. This architecture places all model parameters in a single terabyte-scale memory block, eliminating the need for tensor parallelism or complex partitioning strategies. The result is near-linear performance scaling and dramatically simplified training workflows, with Jais 2 70B requiring significantly less compute than comparable English-optimized models.
Cultural Grounding Beyond Translation
Jais 2's training corpus includes 600 billion Arabic tokens covering Modern Standard Arabic and 17 regional dialects, plus 427,000 Arabic poems with detailed metadata. The model underwent rigorous alignment through supervised fine-tuning, direct preference optimization, and group relative policy optimization to ensure cultural appropriateness. This ground-up design approach contrasts sharply with adapted multilingual models that retrofit Arabic onto English-centric architectures, resulting in superior understanding of dialect variation, politeness norms, and religious reasoning.
Open Access Across Platforms
Both model variants are available as open weights on HuggingFace for developers and researchers. End users can access the system through the web at jaischat.ai and dedicated iOS and Android applications. The multi-platform availability ensures accessibility across the UAE and broader Arabic-speaking world, with the development team encouraging community evaluation and feedback.
Jais 2 represents a strategic shift in global AI development, demonstrating that nations can build frontier-level models without reliance on massive GPU infrastructure or Western tech giants. While companies like OpenAI and Anthropic invest billions in English-first models, Jais 2 achieved state-of-the-art Arabic performance using one-seventh the training data of Llama-3 70B, proving that specialized data curation and purpose-built hardware can match or exceed brute-force scaling approaches.
The project's success validates the sovereign AI thesis gaining traction worldwide, particularly as countries recognize that translated or retrofitted models fail to capture linguistic nuance and cultural context. With Cerebras positioning its wafer-scale systems as an alternative to GPU clusters, Jais 2 serves as both a technical proof point and a political statement: advanced AI capability need not flow exclusively through Silicon Valley's infrastructure.
