Jais 2: A Family of Arabic-Centric Open Large Language Models

Jais 2, developed jointly by MBZUAI, Cerebras, and Inception, is a family of Arabic-centric large language models aimed at advancing Arabic-centric language modeling. The family includes what is reportedly the largest open Arabic-centric LLM trained from scratch at 70B parameters, along with a competitive 8B-parameter variant, both outperforming or matching comparable open models on Arabic benchmarks.

To achieve efficiency, the models use a custom Arabic-centric vocabulary, an optimized architecture, and a compute-efficient training recipe. This allows Jais 2 to train with a substantially smaller token budget than comparable models while maintaining strong Arabic performance and competitive English results.

In evaluation, Jais 2 achieves leading results among open models on OALL2 and AraGen, and performs strongly on culturally grounded Arabic benchmarks such as poetry, religion, cuisine, and dream interpretation, as well as on general tasks like translation and summarization. The models are released on HuggingFace under a commercially permissive license, and Jais 2 70B is also available as a chat app on Web, iOS, and Android.

Running on Cerebras hardware, the 70B model delivers up to 2,000 tokens per second, enabling high-throughput Arabic-centric chat serving. By combining scale, linguistic diversity, cultural fidelity, openness, and speed, Jais 2 provides an open-weight foundation for further research and development in Arabic-centric LLMs.