Inception, Cerebras, And MBZUAI Unveil Jais 2: A Leading Arabic Language Model

Inception, a G42 company, alongside Cerebras Systems and the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), has introduced Jais 2. This is an advanced Arabic Large Language Model (LLM) with 70 billion parameters. It is built on the most comprehensive Arabic-first dataset to date and can be accessed via Inception’s HuggingFace and a web application.

Jais 2 boasts a redesigned architecture and cleaner training data, enhancing its reasoning abilities and fluency in Modern Standard Arabic (MSA) and regional dialects. It also performs well in English. The model is designed to handle real-world linguistic nuances, such as code-switching and informal tones, while maintaining technical and creative capabilities.

The model incorporates a robust safety-first framework supported by instruction-tuning, evaluation, and continuous user feedback. This ensures that Jais 2 remains reliable and adaptable over time. It excels in diverse content areas, including Arabic poetry, culture, and social media language.

Ashish Koshy, CEO of Inception, highlighted the historical limitations of AI in Arabic due to limited datasets. He stated: "The use of Arabic within AI has historically been constrained by limited datasets and fragmented representation. Jais 2 addresses these limitations by bringing unprecedented depth, nuance, and contextual intelligence to the language."

Natalia Vassilieva from Cerebras Systems emphasized the UAE's leadership in developing advanced AI systems. She said: "Jais underscores the UAE’s growing leadership in building advanced, open-weight AI systems. At Cerebras, we’re proud that Jais 2 was both trained and is now served on Cerebras systems."

Cerebras Systems collaborated closely with UAE partners to train Jais 2 using their unique hardware. This partnership allowed them to achieve high-quality results with less computational power compared to similar-sized models.

Impact on Arabic AI Development

Professor Preslav Nakov from MBZUAI noted the challenges faced by Arabic AI development due to scarce high-quality data for training large language models. He remarked: "Arabic has long been underserved in AI development due to limited high-quality data for training large language models."

By significantly enhancing the quality and diversity of Arabic data, Jais 2 provides a foundation reflecting the richness of the Arabic language. This model exemplifies how AI can progress through cultural alignment and open innovation.

The introduction of Jais Chat at jaischat.ai marks a significant step forward for Arabic AI technology.

With inputs from WAM