MBZUAI, Cerebras, and Inception have released Jais 2, a family of Arabic-centric large language models that includes what the researchers claim is the largest open Arabic-centric LLM trained from scratch at 70 billion parameters.

The collaboration produced two model variants: the flagship 70B parameter model and a more compact 8B parameter version. Both models use a custom Arabic-centric vocabulary designed to improve training and inference efficiency for Arabic text processing.

Jais 2 achieved leading performance among evaluated open models on OALL2 and AraGen benchmarks, according to the research paper published on arXiv. The models also demonstrated strong results on culturally specific Arabic tasks including poetry analysis, religious text understanding, cuisine knowledge, and dream interpretation.

Training efficiency breakthrough

The researchers report achieving competitive performance with a substantially smaller token budget compared to similar-scale models. This efficiency gain stems from an optimized architecture and training recipe developed specifically for Arabic language modeling.

Beyond Arabic tasks, Jais 2 maintains competitive English language capabilities across translation and summarization benchmarks. The dual-language competency positions the model for multilingual applications in Arabic-speaking regions.

The 70B model has been deployed as a chat application across web, iOS, and Android platforms. Running on Cerebras hardware, the deployment delivers up to 2,000 tokens per second, enabling high-throughput Arabic conversation serving.

Hugging Face hosts both models under a commercially permissive license, making them available for research and commercial use. The open-weight release follows the trend of major Arabic language model releases becoming publicly accessible.

MBZUAI, the Mohamed bin Zayed University of Artificial Intelligence, led the research collaboration alongside AI chip company Cerebras and investment firm Inception. The partnership represents a significant investment in Arabic-language AI capabilities from the Gulf region.

The models are immediately available for download and deployment, with the chat interface already live across multiple platforms for public testing.