Jais

Jais, launched August 30, 2023, was the world's highest-quality open Arabic language model at release — a 13-billion-parameter bilingual model developed by G42's Inception (now Core42) with the Mohamed bin Zayed University of AI and Cerebras Systems. It was trained on 395 billion Arabic and English tokens on the Condor Galaxy supercomputer, with Arabic deliberately weighted at 33% of the data rather than treated as an afterthought.

The project later expanded into the most extensive open Arabic model family in the region: 20 models from 590M to 70B parameters under the Inception brand, including Llama-2-adapted variants. The license is Apache 2.0, and the original Jais-13B remains a historically important reference point for Arabic NLP. Note that newer Arabic models (including TII's Falcon-H1 Arabic) have since surpassed it on leaderboards.