Photo: Tima Miroshnichenko / Pexels
The Saudi AI company opted for a Chinese open-weights model as the foundation for its Arabic-first national platform, choosing speed over pure homegrown development.
Saudi Arabia’s state-backed AI company Humain released its HUMAIN M3 model on September 2, 2026, positioning it as the backbone of a national AI platform. There’s a catch worth noting: the model everyone initially assumed was built from scratch actually runs on open weights from MiniMax, the Chinese AI firm.
The HUMAIN M3 was originally pitched as a fully sovereign, Arabic-first model. Humain took MiniMax’s open M3 base and trained it on over 1 trillion Arabic tokens, creating something that performs impressively on Arabic benchmarks while relying on external architecture underneath.
What Humain actually built #
The model uses a Mixture-of-Experts (MoE) architecture with approximately 428 billion total parameters and 23 billion active parameters per token, routing each input to the most relevant subset of the model.
On performance, the numbers look solid. HUMAIN M3 averaged around 89.37% across seven Arabic benchmark tests, which places it among the stronger Arabic-language models currently available.
Humain plans to release the model’s weights as open-source and deploy it across Saudi sovereign infrastructure.
The strategic calculus behind the choice #
Humain was founded in May 2025 under the direction of Crown Prince Mohammed bin Salman, backed by the Public Investment Fund. The company fits squarely within Saudi Arabia’s Vision 2030 framework, which has been pouring resources into diversifying the kingdom’s economy beyond oil.
Before HUMAIN M3, the company had promoted its ALLAM 34B model as the flagship Arabic AI system. The pivot to building on MiniMax’s architecture represents a meaningful strategic shift. Rather than spending years developing competitive base models from the ground up, Humain opted to take a high-performing open model and customize it heavily for Arabic.
Humain has also established partnerships with major technology companies including Nvidia, AMD, and Mistral.
Why the MiniMax connection matters #
MiniMax is a Chinese AI company that has been gaining traction by releasing powerful open-weights models. Humain’s decision to build on a Chinese model base is notable given the current geopolitical tensions around AI technology transfer, particularly between the US and China.
Because the model weights are publicly available, Humain isn’t entering a proprietary dependency relationship with a Chinese firm. It’s taking freely available technology and adapting it.
Humain’s broader orchestration strategy aims to aggregate multiple frontier models into a unified deployment system, optimizing for both cost and performance.
What this signals for the AI landscape #
The 1 trillion Arabic tokens used for fine-tuning represent a substantial investment in linguistic and cultural adaptation, even if the underlying architecture wasn’t built in Riyadh.
The model’s performance on Arabic benchmarks also points to a growing market for non-English AI capabilities. Arabic is spoken by over 400 million people, and high-performing Arabic-language models remain relatively scarce compared to English-language alternatives.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our