Saudi Arabia's national AI company HUMAIN unveiled humain-m3, a 428-billion-parameter Arabic language model built on Chinese lab MiniMax's open-weight architecture instead of a Western base. It scores a vendor-reported 89.37% on Arabic benchmarks and is live now in research preview via HUMAIN Node.
Read the full story: HUMAIN's 428B Arabic AI Model Is Built on Chinese Weights →
Transcript
Saudi Arabia just unveiled its flagship Arabic AI model, and the twist is where it came from. HUMAIN built humain-m3 on top of MiniMax-M3, an open-weight model from a Chinese lab, instead of training from scratch or renting a closed American API. It's a 428 billion parameter model, but only about 23 billion parameters activate per token, and HUMAIN layered over a trillion Arabic tokens on top. The company claims 89.37% average across seven Arabic benchmarks, though nobody outside HUMAIN has verified that yet. The real story is strategic. Open Chinese weights are cheap, fast, and you own the result outright. That's a quiet problem for U.S. export controls built around the idea that chips alone keep countries dependent on American models.