HUMAIN Unveils humain-m3 Arabic Frontier Model Built On MiniMax M3 At LEAP 2026

PIF-backed HUMAIN launched humain-m3, a 428B Mixture-of-Experts Arabic language model post-trained on top of China's MiniMax M3, at LEAP 2026 in Riyadh with self-reported top-of-class Arabic benchmark scores.

HUMAIN Unveils humain-m3 Arabic Frontier Model Built On MiniMax M3 At LEAP 2026

Saudi Arabia's state-backed HUMAIN unveiled humain-m3 at LEAP 2026 in Riyadh on September 3, 2026 - a 428-billion-parameter Mixture-of-Experts Arabic language model that the PIF-backed AI company says beat every frontier model it tested across seven public Arabic benchmarks, powered under the hood by Shanghai-based MiniMax's M3.

Arabic Post-Training On A Chinese Open-Weight Foundation

humain-m3 is not trained from scratch. HUMAIN commissioned MiniMax to deliver an Arabic-tuned build of its MiniMax M3 architecture - the 428B MoE with 23B active parameters per token, 1M context window and MiniMax Sparse Attention - then continued pre-training the model on more than one trillion tokens of Arabic-native content. HUMAIN reports the Arabic post-training adds roughly nine percentage points on average over the base MiniMax M3 checkpoint, taking the 80.34% MiniMax reference to a claimed 89.37% across the OALL v2 benchmark suite.

Diagram of MiniMax M3 428B mixture-of-experts architecture that underlies humain-m3

Benchmark Claim With Asterisks

HUMAIN ran the evaluations on its own infrastructure and has not submitted humain-m3 to the independent Open Arabic LLM Leaderboard, so a direct comparison against UAE's Falcon-H1 Arabic 34B (75.36% on OALL) or Qatar's Fanar-2-27B (69.40%) is not yet possible. Open weights are targeted for October 2026 under the MiniMax Community License, which will let researchers reproduce the numbers through the standard OALL pipeline. Compared benchmarks include AlGhafa, ArabicMMLU, Arabic EXAMS, MadinahQA, AraTrust, ALRAGE and Translated MMLU, with HUMAIN placing humain-m3 above GPT-5.6 SOL (87.30%) and Anthropic's Claude Opus 5 (87.34%).

Riyadh's Sovereign-AI Triangle With Beijing And Washington

Running humain-m3 on HUMAIN Node - hosted on the kingdom's own 18,000 NVIDIA GB300 GPUs, AMD's largest inference cluster outside the US, and Groq's 19,000 LPUs - lets Saudi Arabia keep prompt data off MiniMax's Chinese servers even as it builds on Chinese architecture. That threading of needles arrives against a US House Homeland Security investigation into MiniMax and continued case-by-case NVIDIA chip authorizations for HUMAIN under the Commerce Department, while UAE has been elevated to Country Group A:5. Related coverage on HUMAIN's expanding AI stack: HUMAIN + Applied Intuition national-scale autonomous trucking and HUMAIN's strategic investment in Tarjama. MiniMax's Hong Kong-listed shares jumped roughly 10% on the news.

Reporting based on coverage from Tech Times, Unite.AI, PR Newswire and Crypto Briefing.

Category: AI & Technology

Tags: Open Source AI AI Models AI Enterprise AI Partnership AI Foundation Models AI Infrastructure

Related Articles