Jais 2: A Family of Arabic-Centric Open Large Language Models
| Source: arXiv AI
Tags: Jais 2, MBZUAI, Cerebras, Arabic LLM, open-source, multilingual, fine-tuning
MBZUAI, Cerebras, and Inception release Jais 2 — a 70B Arabic-centric open LLM (largest trained from scratch) that tops OALL2 and AraGen benchmarks and runs at 2,000 tokens/second on Cerebras hardware under a commercially permissive license.
Details
Jais 2 is a family of Arabic-centric large language models released jointly by MBZUAI, Cerebras, and Inception. The family includes a 70B model claimed to be the largest open Arabic-centric LLM trained from scratch, alongside a competitive 8B variant. Both are released on HuggingFace under a commercially permissive license.\n\nThe models use a custom Arabic-centric vocabulary designed for training and inference efficiency, letting Jais 2 achieve strong Arabic benchmarks with a substantially smaller token budget than comparable models. On OALL2 and AraGen — the two main Arabic LLM leaderboards — Jais 2 leads among evaluated open models. It also performs well on culturally grounded tasks including poetry, religion, cuisine, and dream interpretation, as well as standard tasks like translation and summarization.\n\nThe 70B chat variant is deployed via a web app, iOS, and Android app running on Cerebras hardware, delivering up to 2,000 tokens per second. English performance is described as competitive but secondary to the Arabic focus. For teams building Arabic-language applications, Jais 2 is now the clear open-weight baseline to evaluate against.