Saudi AI Firm HUMAIN Unveils 428B-Parameter Arabic Model

HUMAIN, the artificial intelligence company backed by Saudi Arabia's Public Investment Fund (PIF), unveiled a new frontier Arabic-language model called humain-m3 on September 3 at the LEAP technology conference in Riyadh, according to a company press release distributed by PR Newswire. The model, commissioned by HUMAIN and delivered by Chinese AI lab MiniMax, is a 428-billion-parameter mixture-of-experts system built on the MiniMax-M3 architecture, further pre-trained on more than one trillion tokens of Arabic-native text. It became available the same day in a research-preview tier on HUMAIN Node, the company's platform for developers, researchers and enterprises, according to a detailed report by AI news outlet Unite.AI, which reviewed the benchmark results HUMAIN published on its own model page.

According to figures HUMAIN published on its Node model page and relayed by Unite.AI, the previewed humain-m3 checkpoint scored an average of 89.37% across seven public Arabic benchmarks — ahead of GPT-5.6 SOL at 87.30% and Anthropic's Opus 5 at 87.34%, and nine points above the MiniMax-M3 reference checkpoint it is built on. The model reportedly leads five of the seven benchmarks against the strongest systems HUMAIN tested, including a score of 90.70% on ArabicMMLU, a broad-knowledge test, and 97.53% on AraTrust, a benchmark for factual reliability. HUMAIN said the comparisons reflect its own internal evaluation of the preview checkpoint; the scores have not been submitted to the independently run Open Arabic LLM Leaderboard, so outside verification does not yet exist.

An Arabic model built on Chinese foundations

The model's architecture points to a strategy already visible elsewhere in the region: building sovereign AI products on top of open-weight foundations released by Chinese labs, rather than starting from scratch or licensing exclusively from US developers. MiniMax, the Shanghai-based lab that delivered the underlying M3 architecture, describes itself as serving more than 300 million users and two million enterprises and developers worldwide. HUMAIN frames the project as reinforcing "the Arabic AI ecosystem while contributing to the global development of frontier intelligence," and its chief executive, Tareq Amin, said in the press release that "Arabic is spoken by hundreds of millions of people, yet it remains significantly underrepresented at the frontier of artificial intelligence."

Unite.AI's review of the HUMAIN Node documentation describes humain-m3 as a natively multimodal system, trained jointly on text, image and video from the first step, that activates 23 billion of its 428 billion parameters per token and offers three "thinking" modes — always-on, adaptive and off. HUMAIN said the system is built for long-horizon agent workflows, including frontier tool use and computer-use tasks, a design goal shared by the broader shift toward agentic systems capable of planning and executing multi-step tasks rather than only answering single prompts.

• Model size: 428 billion parameters, 23 billion active per token (mixture-of-experts)
• Training data: more than 1 trillion tokens of Arabic-native content added to the MiniMax-M3 base
• Average score across 7 public Arabic benchmarks: 89.37%, versus 87.30% for GPT-5.6 SOL and 87.34% for Opus 5
• Access: research preview via HUMAIN Node from September 3; open-weight release targeted for the following month under the MiniMax Community License

What it changes beyond Saudi Arabia's own market

HUMAIN, launched by the PIF in 2025 as Saudi Arabia's national AI champion, has already built the ALLaM family of Arabic models and struck partnerships with Microsoft, Nvidia, AMD, Qualcomm and Groq, hosting what it says are now more than 100 model providers on HUMAIN Node. Positioning humain-m3 alongside frontier Western systems, rather than only regional Arabic tools, signals an ambition that reaches past the Gulf: a model that outperforms GPT-5.6 SOL and Opus 5 on Arabic-language tasks matters to any company — in the Levant, North Africa or the Arabic-speaking diaspora — building customer support, education or government services in Arabic, a market where general-purpose Western models have historically underperformed on dialect and cultural nuance. It also offers a concrete example, rather than a hypothetical one, for AI groups elsewhere weighing whether to build national-language products on Chinese open-weight foundations instead of waiting for a Western equivalent.