Inception, Cerebras and MBZUAI release Jais 2 Arabic LLM

Open-weight model trained on largest Arabic dataset ever assembled

The Jais Arabic-centric LLM is named after Jebal Jais in the north of the UAE (Image credit: MEAIN via Google Gemini)

#UAE #LLMs - G42 group company Inception, California-based AI infrastructure provider Cerebras and the Institute of Foundation Models of Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) have released Jais 2, billing it as the world’s leading open-weight Arabic large language model. The 70-billion parameter model was built from the ground up and trained on the largest Arabic-first dataset ever assembled, delivering enhanced reasoning capabilities, fluency across Modern Standard Arabic (MSA) and regional dialects, coupled with strong English performance. Jais-2-8B-Chat and Jais-2-70B-Chat are now available for download via AI community side Hugging Face, while there is a separate Jais 2 web chat site. Jais 2 incorporates a comprehensive safety framework and demonstrates proficiency across Arabic poetry, cultural content and social media communication.

SO WHAT? - Jais 2 is arguably the most significant advancement in the Jais series of Arabic-centric large language models. By training on the richest Arabic-first dataset to-date from scratch, with redesigned model architecture, Jais 2 delivers improvements in understanding real-world linguistic behaviour including code-switching, informal tone and regional dialect variations. According to Inception, it also provides stronger reasoning. The open-weight approach enables organisations across the Arab world to access state-of-the-art Arabic AI capabilities .

Here are some key points about the new Jais 2 release:

• G42 group company Inception, Ai chip developer Cerebras and the Institute of Foundation Models of Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) have announced Jais 2, the next generation of their Arabic large language model with 70 billion parameters.

• The open-weight model was built from the ground up and trained on the largest Arabic-first dataset ever assembled, featuring redesigned architecture and cleaner training data to deliver stronger reasoning capabilities and greater fluency across Modern Standard Arabic and regional dialects.

• The Jais-2-8B-Chat and Jais-2-70B-Chat models are available for download via open-source community Hugging Face, providing access to researchers, developers and organisations seeking to build Arabic-centric AI applications and services. Meanwhile, a Jais Chat web application is online to encourage public engagement. (links can be found at the end of this article)

• Jais 2 is engineered to navigate real-world linguistic behaviour including code-switching between Arabic and English, informal tone and colloquial expressions, whilst maintaining advanced technical and creative capabilities across domains including Arabic poetry, cultural content and social media communication.

• The model integrates a comprehensive safety-first framework supported by instruction-tuning, evaluation protocols and continuous user feedback mechanisms, ensuring reliability and adaptability over time whilst maintaining cultural alignment and appropriate content generation for Arabic-speaking audiences.

• Cerebras trained Jais 2 on its Wafer Scale Engine infrastructure, achieving state-of-the-art quality using a fraction of the compute resources typically required for similar-sized models, demonstrating efficiency advantages in developing large-scale Arabic language models.

• Within the G42 ecosystem, Inception functions as the core intelligence layer transforming data and compute infrastructure into applied AI solutions, having previously developed bilingual large language models for Arabic (Jais), Hindi (Nanda) and Kazakh (Sherkala) to ensure language accessibility.

ZOOM OUT - The Jais 2 release builds upon Inception’s substantial 2024 momentum in Arabic AI development. In the second half of last year, the company released a comprehensive family of 20 open-source Arabic-centric Jais large language models, ranging from 590 million to 70 billion parameters and trained on up to 1.6 trillion tokens of Arabic, English and code data. This represented the largest single release of production-ready models in the Middle East and North Africa region. Jais 2 now represents the next evolutionary step, delivering enhanced capabilities whilst maintaining the open-weight philosophy that has enabled widespread access across Arabic-speaking markets.

[Written and edited with the assistance of AI]

LINKS

• Jais 2 Chat (website)

• Jais 2 Family space (Hugging Face)

• Jais 2 Android app (Play Store)

• Jais 2 iOS app (App Store)

• Jais 2 technical paper (pending)

Read more about Jais Arabic-centric LLM:

• Inception launches new JAIS Chat mobile app (Middle East AI News)

• G42 launches 20 open-source Jais AI models (Middle East AI News)

• Will GenAI champion the Arabic language? (Middle East AI News)

Read about other G42 language models:

• Inception & MBZUAI unveil Kazakh LLM (Middle East AI News)

• MBZUAI open-sources NANDA LLM (Middle East AI News)

• G42 launches new Hindi LLM in Mumbai (Middle East AI News)