Jais 2 family introduces 70B and 8B Arabic-centric open LLMs with culturally grounded benchmarks

MBZUAI, Cerebras, and Inception jointly released Jais 2, a family of Arabic-centric open large language models designed to advance Arabic-centric language modeling and culturally grounded evaluation. The family includes a 70B-parameter model and an 8B-parameter variant, with the 70B model described as the largest open Arabic-centric LLM trained from scratch to date.

The models use a custom Arabic-centric vocabulary and an optimized architecture and training recipe to improve compute efficiency and reduce token budget requirements. Despite a smaller token budget than comparable models, Jais 2 achieves strong Arabic performance on the benchmarks evaluated in the report and competitive English results.

On the OALL2 and AraGen benchmarks, Jais 2 obtains leading results among the evaluated open models. The models also perform strongly on culturally grounded Arabic benchmarks covering poetry, religion, cuisine, and dream interpretation, as well as general tasks such as translation and summarization.

Jais 2 70B is released under a commercially permissive license on Hugging Face and as a chat application for web, iOS, and Android. The 70B model runs on Cerebras hardware and is reported to deliver up to 2,000 tokens per second in the deployment setting, enabling high-throughput Arabic-centric chat serving.