MiniMax's Open-Weight M3 Becomes the Base for Saudi Arabia's Sovereign HUMAIN-M3
The AMW Read
A Chinese open-weight model becoming the training base for a PIF-backed national model is a meaningful new sovereign-AI adoption pattern for the segment, but it updates rather than overturns the existing player map.
MiniMax's Open-Weight M3 Becomes the Base for Saudi Arabia's Sovereign HUMAIN-M3
At LEAP 2026 on September 3, HUMAIN — the Saudi PIF-backed AI company chaired by Crown Prince Mohammed bin Salman — launched HUMAIN-M3, an Arabic-focused model built on MiniMax's open-weight M3. It keeps M3's mixture-of-experts architecture (428B total parameters, ~23B active per token) and trained further on over 1 trillion Arabic-native tokens. Across seven Arabic benchmarks it averaged 89.37%, topping five of seven and beating GPT-5.6 Sol (87.30%) and Claude Opus 5 (87.34%). The turnaround from M3's release to HUMAIN-M3 was about three months.
The deal marks a broader shift: open-weight Chinese models becoming the base layer for sovereign AI. CNAS's Sovereign AI Index counts 184 government-backed sovereign AI projects across 67 countries as of mid-2026, with 41 launched in H1 alone — more than all of 2024. Governments cite data residency, deployment control, and service continuity, sharpened by June's export-control suspension of Fable 5. Open weights let a national project start near the frontier and fine-tune locally rather than build from scratch. MiniMax is also stacking infrastructure around M3: SambaNova's SN50 chips posted the fastest measured inference for the model per Artificial Analysis, and Nebius made M3 the first open-weight model with an exclusive slot on its Token Factory platform.
Per the AI Market Watch index, MiniMax's news volume held roughly steady over the past 90 days (20 items versus 18 prior; name-matched pipeline coverage only) heading into this deal. For builders and investors: CN open-weight labs are shifting revenue beyond token-metered API access toward deployment, fine-tuning, and support contracts with state-backed buyers, and labs pairing releases with chip- and cloud-hosting partnerships build switching costs that make displacement harder once a customer is locked in.
