HUMAIN Unveils Arabic-Focused humain-m3, a 428B-Parameter MoE Built by China's MiniMax

Saudi AI company HUMAIN announced humain-m3, a large language model specialized for Arabic, on September 3, 2026, at LEAP in Riyadh. The model carries 428 billion parameters, and the company that actually built it is China's MiniMax. Choosing a Chinese foundation as the base for a national frontier model in Arabic is what stands out here.

Deciding not to build from scratch

humain-m3 sits on the MiniMax-M3 lineage that MiniMax publishes openly. HUMAIN commissioned the work and MiniMax delivered it, with more than one trillion tokens of Arabic-native content added through further pre-training. Both the base weights and the architecture originate in China.

National-scale AI investment usually implies assembling a foundation model entirely in house. HUMAIN went around that. Starting from a strong publicly available base, it concentrated its resources on native-language data and post-training. The distance to the frontier gets closed through procurement rather than through time.

HUMAIN is a company under Saudi Arabia's Public Investment Fund (PIF), and it positions itself as a vertically integrated stack spanning data centers, cloud infrastructure, models, and business solutions. It already has its own Arabic model family, ALLAM. This launch runs alongside that effort while widening the lane for outside technology.

An average of 89.37 percent across seven benchmarks

In HUMAIN's own evaluation, the previewed humain-m3 checkpoint averaged 89.37 percent across seven public Arabic benchmarks. The comparisons the company cited were GPT-5.6 SOL at 87.30 percent, Opus 5 at 87.34 percent, and the MiniMax M3 reference checkpoint at 80.34 percent. HUMAIN said the model leads on five of the seven.

The breakdown runs as follows: 86.45 percent on AlGhafa for core Arabic understanding, 90.70 percent on ArabicMMLU for broad native-Arabic knowledge, 67.67 percent on Arabic EXAMS for academic examinations, 95.44 percent on MadinahQA for language proficiency, 97.53 percent on AraTrust for truthfulness, 94.63 percent on ALRAGE for retrieval-augmented generation, and 93.20 percent on Translated MMLU.

The gap against the underlying MiniMax M3 is the number worth watching. Nine points on average were added, which is Arabic post-training showing up directly in the results. Put another way, those nine points are HUMAIN's contribution.

All of these figures come from HUMAIN measuring its own previewed checkpoint. No third-party replication has appeared yet, and Arabic EXAMS at 67.67 percent sits clearly below the other categories, so treating this as pending external verification is the reasonable reading.

Designed around agents

The architecture is a mixture of experts (MoE, where only part of the expert networks fire for a given input). Of the 428 billion total parameters, 23 billion activate per token. HUMAIN describes it as a natively multimodal MoE engineered for agents.

Text, images, and video were trained jointly from the first step, and the model is said to handle long-video understanding and native screen operation. The claim is that it can carry out long-horizon agent work, including tool use and computer use, in both Arabic and English.

Three thinking modes are available: always-on, adaptive, and off. Whether to spend time on reasoning becomes a per-use-case setting, which lets response speed and cost be tuned in production.

Available on HUMAIN Node, with weights targeted for next month

humain-m3 is available as a research preview on HUMAIN Node, the company's model access platform. It can be tried in a no-code playground or called through an OpenAI-compatible API endpoint.

Access is tiered. Creating an account opens the ability to request access. The limited-preview tier applies a Saudi alignment guardrail, with thinking and streaming turned off and somewhat higher latency. The research-preview tier provides the full checkpoint with thinking and streaming enabled and lower latency.

HUMAIN frames the preview period as a way to evaluate capability, safety, and alignment across Arabic dialects together with early users before general availability, with feedback shaping the next checkpoints.

HUMAIN Node is said to already carry more than 100 models, with a single API key and bill across them, spending limits by team and project, and a choice of global, in-Kingdom, or sovereign hosting. humain-m3 is the first model opened to preview users on that platform.

The company also addressed the weights. It expects to release them under the MiniMax Community License once safety training and alignment are complete, currently targeting next month. A foundation that started out open would then return to open form, this time with heavy Arabic training layered on.

Summary

humain-m3 is an attempt to fill a frontier-model gap for the Arabic-speaking world, and its distinguishing feature is the choice of an open Chinese base as the means. HUMAIN CEO Tareq Amin said Arabic is spoken by hundreds of millions of people yet remains significantly underrepresented at the AI frontier, framing the model as an investment in changing that. Because the benchmark figures are self-reported, judging the model's real strength will have to wait for outside verification after the weight release planned for next month.