News / Model Watch · MiniMax going global
Saudi Arabia's "100% Saudi Model" Runs on MiniMax M3: Chinese Open Weights Just Became a National Base Layer
HUMAIN, Saudi Arabia's national AI company, released HUMAIN-M3 — a 428B-parameter MoE marketed as a homegrown Saudi model. Multi-source reporting confirms what the "纯血" (pure-blood) marketing glossed over: the base is MiniMax M3, the open-source model from China's MiniMax, post-trained on 1T+ tokens of native Arabic data. It's the first large Arabic-language model on a Chinese open-weight foundation — and the clearest signal yet that Chinese open weights are becoming the infrastructure layer for other countries' sovereign AI.
HUMAIN-M3 = MiniMax M3 open weights + Saudi post-training: over 1 trillion native Arabic tokens driving Arabic understanding, cultural adaptation, and agent skills (tool calling, computer operation). Live as a research preview on HUMAIN NODE, with announced plans to open weights and support sovereign deployment. For the Chinese-model ecosystem this is a first: an open-weight foundation powering another nation's flagship — commentary calls it "AI-era overseas infrastructure", with Qwen, DeepSeek and GLM named as likely future national-base candidates.
What's confirmed across sources
- The release: HUMAIN-M3, a 428B-parameter MoE from Saudi Arabia's HUMAIN, positioned as a national model; available as research preview on HUMAIN NODE.
- The base: MiniMax M3 (稀宇科技 open-source release) — confirmed by 解放日报 ("基于MiniMax M3的首个阿拉伯语大模型"), 财联社/东方财富, and 36kr's own teardown of the "100% Saudi" claim.
- The adaptation: 1T+ native Arabic tokens of post-training; focus areas are Arabic understanding/generation, local cultural adaptation, and agent capabilities including tool calling and computer operation.
- The roadmap: announced plans to open the model weights and support sovereign deployment for Saudi domestic AI capability.
- The framing: Chinese coverage treats the marketing-versus-provenance gap as the story ("惊天大瓜"), while the substantive read is a new export model for Chinese open weights — national-base licensing rather than API sales.
Why this pattern matters
- Open weights are the product. MiniMax's revenue path here isn't tokens served — it's the ecosystem gravity of M3 being chosen as a national foundation. The Design launch and this story together show MiniMax running a two-track commercial strategy (products + base-model adoption).
- Sovereignty gets redefined. "100% Saudi" now means: our data, our alignment, our deployment — on someone else's frontier base. Expect more governments to make the same trade; expect the marketing-vs-provenance gap to keep being news.
- The candidate bench is Chinese. Commentary names MiniMax, Qwen, DeepSeek and GLM as the plausible national-base suppliers — all four have 2026 open-weight releases at or near frontier capability. This is the open-weights-as-leverage thesis playing out at state scale.
- For developers: none of this changes MiniMax M3's own API availability or pricing today — but base-model adoption at this scale is a leading indicator of long-term maintenance investment in the open releases.
Primary sources
FAQ (2026)
What is HUMAIN-M3?
Saudi national AI company HUMAIN's 428B MoE, marketed as "100% Saudi" — built on MiniMax M3 open weights per multi-source reporting.
How was it adapted?
1T+ native Arabic tokens of post-training: Arabic understanding, cultural adaptation, agent skills (tool calling, computer operation).
Where to try it?
Research preview on HUMAIN NODE; weight release and sovereign-deployment support announced as roadmap items.
Why "纯血" was contested?
The sovereignty claim covers data/alignment/control; the foundation is MiniMax M3. The gap is the story, per 36kr's teardown.
Who's next?
Commentary names MiniMax, Qwen, DeepSeek and GLM as likely national-base suppliers — the open-weight frontier bench.
Does it change MiniMax APIs?
No rate or availability change today — it's a leading indicator of long-term investment in the open M3 line.