HUMAIN unveiled humain-m3 on September 3, 2026, at LEAP in Riyadh. The Public Investment Fund company commissioned the model. MiniMax built it. A research preview is live today on HUMAIN Node. Open weights are planned for next month under the MiniMax Community License, after safety training and alignment.

humain-m3 is a 428-billion-parameter mixture-of-experts model further pre-trained on more than one trillion tokens of Arabic-native content, on top of the MiniMax-M3 lineage. HUMAIN said it posted the highest average score among the frontier models it evaluated on seven public Arabic benchmarks.

The company’s own comparison table, released with the announcement, puts that average at 89.37 percent, ahead of GPT-5.6 SOL at 87.30 percent, Opus 5 at 87.34 percent, and an M3 reference at 80.34 percent. humain-m3 leads on AlGhafa (86.45 percent), ArabicMMLU (90.70 percent), Arabic EXAMS (67.67 percent), MadinahQA (95.44 percent), and AraTrust (97.53 percent). It trails slightly on ALRAGE retrieval and on a translated MMLU set.

A Node preview, then promised weights

HUMAIN Node is the access path. The company describes it as a place for developers, researchers, and enterprises to reach models and inference, including HUMAIN’s own line and third-party systems. humain-m3 starts there as an evaluation preview so testers can send feedback before the weight drop.

Chief executive Tareq Amin tied the release to a gap the company has been working since its ALLAM Arabic models.

“Arabic is spoken by hundreds of millions of people, yet it remains significantly underrepresented at the frontier of artificial intelligence. With humain-m3, we are investing in changing that. And through HUMAIN Node, we are making that intelligence accessible so developers, researchers and innovators can experiment with it, build on it and create the next generation of Arabic AI experiences.” Tareq Amin, CEO, HUMAIN

The announcement sits next to other LEAP deals the same week, including a HUMAIN collaboration with Reflection AI and an expanded Microsoft partnership. Those are separate products. humain-m3 is the language-model piece: a Chinese lab’s MoE stack, further trained on Arabic text, sold through a Saudi inference door.

Decoded Take

HUMAIN is buying Arabic fluency the way it has been buying compute: write the check, put the model on Node, promise weights later. A 428 billion parameter MoE that wins company-run Arabic averages is useful if ministries and banks can actually host it, or at least call it without sending prompts out of the Kingdom. The MiniMax Community License and the “next month” safety delay are the constraints. Watch whether the open-weight drop lands on schedule, whether independent Arabic benches confirm the 89.37 percent average, and whether Node traffic looks like research curiosity or production replacement for ALLAM and imported frontier APIs.