From Model Export to Sovereign AI: Saudi Arabia's HUMAIN Builds the First Arabic Large Model Based on MiniMax M3

On September 3, it was announced that the Saudi Public Investment Fund (PIF) has launched the Arabic large model HUMAIN M3 through its artificial intelligence company HUMAIN. This model is based on the open-source flagship model MiniMax M3 from the Chinese AI company MiniMax (00100) and has been fine-tuned using over one trillion tokens of Arabic data. Currently, HUMAIN M3 is available for research preview on the HUMAIN Node platform. Unlike past Chinese large models that primarily entered overseas markets through APIs or application products, this collaboration sees MiniMax M3 serving as the foundational model for the development of local models abroad. HUMAIN is training Arabic language and localization capabilities based on MiniMax M3 and plans to further open its model weights and support sovereign deployment, providing fundamental technical support for building local AI capabilities in Saudi Arabia. HUMAIN is a crucial player in Saudi Arabias effort to develop local AI capabilities. The company is chaired by Crown Prince Mohammed bin Salman, with business operations encompassing data centers, cloud infrastructure, models, and applications, all of which play a strategic role in advancing Saudi Arabia's Vision 2030 and enhancing local technology and innovation capabilities. According to evaluation results released by HUMAIN, HUMAIN M3 achieved an equal-weighted average score of 89.37% across seven publicly available Arabic benchmark tests, attaining the highest score in five of those tests. The evaluations cover tasks such as Arabic language comprehension, knowledge, linguistic capabilities, authenticity, and retrieval-augmented generation. HUMAIN M3 utilizes a mixture of experts (MoE) architecture with 428 billion parameters, activating 23 billion parameters for each token. HUMAIN has conducted over one trillion tokens of specialized Arabic training on the base model, focusing on enhancing Arabic comprehension and generation, adapting to local culture, and capabilities like agent interaction, tool invocation, and computer operations. HUMAIN noted that hundreds of millions of people worldwide speak Arabic, yet the coverage of this language in cutting-edge AI remains insufficient. With HUMAIN M3 and HUMAIN Node, the company aims to enable developers to build AI applications targeting Arabic-speaking users. MiniMax M3, which was released and open-sourced in June this year, is the flagship model developed by MiniMax for coding, agents, and native multimodal scenarios. From the early stages of training, the model underwent multimodal mixed training involving text, images, and videos, supporting a context of up to one million tokens. The large-scale language and localization training HUMAIN is conducting based on M3 also reflects the deep value of open models within the global AI ecosystem: developers, enterprises, and even national-level AI platforms can further train proprietary models tailored to their languages, cultures, and application scenarios based on foundational models. This trend is rapidly emerging within the global open-source AI ecosystem. The report "State of Open Source on Hugging Face: Spring 2026" released by Hugging Face this year indicates that Chinese-developed models accounted for 41% of platform model downloads in the past year, surpassing the United States for the first time. Chinese open-source models are evolving from direct usage by overseas developers to more in-depth processes like model retraining, local deployment, and industry customization. For Chinese foundational model companies, the meaning of "model going overseas" is changing: beyond simply providing APIs and products, open models themselves are becoming the technical foundation for global developers and institutions to build the next layer of AI capabilities.