MBZUAI, Cerebras, and Inception Release Jais 2, a Family of Arabic-Centric Open LLMs

The Jais 2 family comprises large language models focused on Arabic, collaboratively developed by MBZUAI, Cerebras, and Inception to enhance Arabic language processing. This collection features the largest open Arabic-centric LLM, trained from the ground up with 70 billion parameters, alongside a robust 8 billion-parameter variant among assessed open models. A specialized Arabic vocabulary facilitates effective training and inference, while an optimized architecture ensures efficient computational training. With a notably reduced token budget compared to similar models, Jais 2 demonstrates impressive Arabic performance on the benchmarks highlighted in the report, as well as competitive results in English. The models excel in OALL2 and AraGen evaluations and show strong performance on various cultural benchmarks. The findings are documented in a report available on arXiv under identifier 2608.13580.