On September 3rd, HUMAIN, an artificial intelligence company under Saudi Arabia's Public Investment Fund (PIF), officially launched its Arabic large language model, HUMAIN M3. The model is built upon the open-source flagship model MiniMax M3 from Chinese AI firm MINIMAX-W, and has undergone post-training using more than one trillion tokens of Arabic data. Currently, HUMAIN M3 is available for research preview through the HUMAIN Node platform.
Unlike previous instances where Chinese large models entered overseas markets primarily through APIs or application products, this collaboration marks a shift as MiniMax M3 has directly become the foundational model for developing a native model overseas. HUMAIN has trained on MiniMax M3 to enhance Arabic language and localization capabilities, and plans to further open model weights and support sovereign deployment, providing the underlying technical support for building Saudi Arabia's indigenous AI capabilities.
HUMAIN is a key player in Saudi Arabia's efforts to establish domestic AI capabilities. The company, chaired by Crown Prince Mohammed bin Salman, operates across data centers, cloud infrastructure, models, and applications, and is tasked with advancing the nation's "Vision 2030" and developing local technology and innovation capacity.
According to official evaluation results released by HUMAIN, the HUMAIN M3 model achieved an equally weighted average score of 89.37% across seven public Arabic benchmark tests, securing the highest score in five of them. The evaluations covered tasks such as Arabic comprehension, knowledge, language proficiency, factuality, and retrieval-augmented generation.
HUMAIN M3 adopts a Mixture-of-Experts (MoE) architecture with 428 billion parameters, activating 23 billion parameters per token. HUMAIN conducted specialized training exceeding one trillion tokens of Arabic data on the base model, with a focus on strengthening Arabic understanding and generation, local cultural adaptation, as well as capabilities in agents, tool calling, and computer operation. HUMAIN stated that hundreds of millions of people worldwide speak Arabic, yet this language remains under-represented in frontier artificial intelligence. Through HUMAIN M3 and HUMAIN Node, the company hopes developers can build AI applications further tailored to Arabic-speaking users.
MiniMax M3 was released and open-sourced in June this year, serving as MiniMax's flagship model designed for coding, agent applications, and native multimodal scenarios. From the early stages of training, the model has engaged in mixed multimodal training involving text, images, and video, supporting million-token-level context. HUMAIN's large-scale language and regional training based on M3 reflects the deeper value of open models in the global AI ecosystem: developers, enterprises, and even national-level AI platforms can further train specialized models tailored to their own languages, cultures, and application scenarios on top of a base model. This trend is accelerating across the global open-source AI ecosystem.
The "State of Open Source on Hugging Face: Spring 2026" report released by Hugging Face this year shows that over the past year, models developed by Chinese entities accounted for 41% of all model downloads on the platform, surpassing the United States for the first time. Chinese open-source models are extending beyond direct use by overseas developers into deeper areas such as model retraining, local deployment, and industry-specific customization. For Chinese foundational model companies, the meaning of "model exports" is transforming: beyond delivering APIs and products, open-sourcing models themselves is becoming the technological basis for developers and institutions worldwide to build the next layer of AI capabilities.