What was announced
HUMAIN, the AI company owned by Saudi Arabia’s Public Investment Fund, unveiled humain-m3 on 3 September as a research and evaluation preview on its own platform, HUMAIN Node.
The model is a mixture-of-experts system with 428 billion total parameters and about 23 billion active per token. HUMAIN says it was further pre-trained on more than one trillion tokens of Arabic-native content, and that it achieved the highest average performance across seven public Arabic benchmarks among the frontier models the company evaluated. That is HUMAIN’s own evaluation, not an independent one.
“Arabic is spoken by hundreds of millions of people, yet it remains significantly underrepresented at the frontier of artificial intelligence,” said Tareq Amin, HUMAIN’s chief executive.
The part in the headline
The release describes humain-m3 as “commissioned by HUMAIN and delivered by MiniMax”, the Shanghai lab. It is built on the MiniMax-M3 lineage — the same 428-billion-parameter architecture, which is how observers matched the two.
Bloomberg reported the model as based on China’s MiniMax. That framing sits awkwardly beside the sovereign-AI language that has surrounded HUMAIN since the PIF launched it, and it is the reason this release is worth reporting at all.
The qualifier matters, though. MiniMax-M3’s weights are openly available, so building on them is not the same as depending on a Chinese vendor’s API or its permission. HUMAIN owns the resulting weights and the Arabic training run. What it does not own is the base.
Sovereignty, redefined downward
This is the practical shape sovereign AI is taking outside the United States and China. Training a 428-billion-parameter model from scratch costs a great deal of compute and several years of accumulated practice. Post-training someone else’s open base on a trillion tokens of your own language costs a fraction of both and gets a usable national model into the field this year.
What it does not get you is control over the pretraining data, the architecture choices, or the safety behaviour baked in before you started. For a state model intended to serve government and public services, those are not small omissions — and they are difficult to audit from the outside.
When the weights arrive
HUMAIN says it expects to release the model weights under the MiniMax Community License once safety training and alignment are complete, currently targeted for next month. Until then, humain-m3 is available only as a preview on HUMAIN Node.
What to watch next
Two things. Whether the weights actually ship next month and under which terms, since a community licence is not the same as Apache 2.0. And whether independent Arabic-language evaluation confirms the benchmark claim — Arabic evaluation is thin enough that a vendor average across seven tests can move a long way on test selection alone.
