HUMAIN unveils Arabic AI model built on China's MiniMax foundation

Saudi Arabia’s state-backed AI company, HUMAIN, has unveiled humain-m3, a large language model specifically optimized for the Arabic language. Rather than developing a foundation model from scratch, HUMAIN utilized the open-weight MiniMax-M3 architecture from China as a base, performing extensive post-training on over one trillion Arabic tokens.

The model employs a Mixture-of-Experts (MoE) architecture with 428 billion total parameters, activating approximately 23 billion per token. This approach allows the model to understand various Arabic dialects, local contexts, and specific enterprise or public service scenarios. In research previews, humain-m3 reportedly achieved an average score of approximately 89.37% across seven public Arabic benchmark tests.

This development highlights a shift in the “sovereign AI” landscape, where nations focus on controlling data, deployment, and cultural alignment rather than solely owning the underlying algorithms. HUMAIN, supported by the Saudi Public Investment Fund (PIF), is integrating this technology with significant infrastructure investments, including AI data centers in NEOM’s Oxagon, to ensure sensitive data remains within the country.