Jais 2: A Blueprint for Sovereign AI (www.cerebras.ai)

G42, in collaboration with the Institute of Foundation Models at MBZUAI and Cerebras Systems, has announced Jais-2, a new family of Arabic-centric Large Language Models (LLMs) designed to provide robust and culturally aligned AI capabilities for Arabic-speaking nations. This initiative aims to bridge the gap between existing Arabic models, which often lack depth and diversity, and the more developed English models, by utilizing 2.6 trillion tokens for training and achieving state-of-the-art performance in various Arabic applications. Jais 2 models operate with impressive speed, processing up to 2,000 tokens per second, making them among the fastest in the world. The significance of Jais-2 extends beyond technical performance; it serves as a blueprint for sovereign AI development that emphasizes cultural relevance and efficiency. By leveraging Cerebras wafer-scale architectures, Jais 2 reduces the complexity and costs associated with traditional GPU clusters, paving the way for other nations to create advanced AI systems tailored to local needs. With its open-weight models available on HuggingFace, Jais 2 encourages further exploration and research in the AI/ML community, demonstrating that nations can achieve frontline AI capabilities grounded in their unique linguistic and cultural contexts.