HUMAIN, the Saudi AI company backed by the Public Investment Fund, unveiled humain-m3 at LEAP in Riyadh on 3 September: a 428 billion parameter mixture of experts model with 23 billion active parameters per token, built on the MiniMax-M3 lineage, commissioned by HUMAIN and delivered by the Chinese lab MiniMax, and further pre trained on more than one trillion tokens of Arabic content.1 HUMAIN via PR Newswire 2026-09-03 428B mixture of experts on the MiniMax-M3 lineage; commissioned by HUMAIN, delivered by MiniMax; more than one trillion Arabic tokens; highest average on seven Arabic benchmarks; research preview on HUMAIN Node; weights targeted for next month under the MiniMax Community License; LEAP 2026 Riyadh. Open source 2 Crypto Briefing 2026-09-02 23 billion active parameters; 89.37 percent average; built on the open MiniMax-M3 base; HUMAIN founded May 2025 under PIF; ALLAM 34B as prior flagship; Nvidia, AMD and Mistral partnerships; pragmatic pivot framing. Open source HUMAIN's own evaluation puts it at an 89.37 percent average across seven public Arabic benchmarks, ahead of GPT-5.6 SOL at 87.30 and Opus 5 at 87.34.3 Unite.AI 2026-09-03 Per benchmark scores; comparison to GPT-5.6 SOL 87.30, Opus 5 87.34, MiniMax M3 80.34; two preview tiers including a Saudi alignment guardrail tier; more than 100 models on HUMAIN Node. Open source It is available in research preview through HUMAIN Node, with weights promised next month under the MiniMax Community License.1 HUMAIN via PR Newswire 2026-09-03 428B mixture of experts on the MiniMax-M3 lineage; commissioned by HUMAIN, delivered by MiniMax; more than one trillion Arabic tokens; highest average on seven Arabic benchmarks; research preview on HUMAIN Node; weights targeted for next month under the MiniMax Community License; LEAP 2026 Riyadh. Open source Our assessment, with high confidence, is that the important fact is the supplier: a Gulf state's flagship sovereign model is a Chinese open weight base. With moderate confidence, we assess the benchmark lead as real but narrow and self reported. With moderate confidence, we assess this as the template for sovereign AI outside the US and China: buy the base, own the data and the deployment, call the result national.

What was built and by whom

The lineage is explicit. HUMAIN commissioned the model and MiniMax delivered it, starting from MiniMax's own M3 base and continuing pre training on Arabic data.1 HUMAIN via PR Newswire 2026-09-03 428B mixture of experts on the MiniMax-M3 lineage; commissioned by HUMAIN, delivered by MiniMax; more than one trillion Arabic tokens; highest average on seven Arabic benchmarks; research preview on HUMAIN Node; weights targeted for next month under the MiniMax Community License; LEAP 2026 Riyadh. Open source Crypto Briefing frames this as a pragmatic pivot: HUMAIN's earlier flagship was ALLAM, a 34 billion parameter Arabic model, and the company chose speed over building a frontier base from scratch.2 Crypto Briefing 2026-09-02 23 billion active parameters; 89.37 percent average; built on the open MiniMax-M3 base; HUMAIN founded May 2025 under PIF; ALLAM 34B as prior flagship; Nvidia, AMD and Mistral partnerships; pragmatic pivot framing. Open source The result is natively multimodal with three thinking modes, and it ships through HUMAIN Node, a platform hosting more than 100 models behind one OpenAI compatible API.3 Unite.AI 2026-09-03 Per benchmark scores; comparison to GPT-5.6 SOL 87.30, Opus 5 87.34, MiniMax M3 80.34; two preview tiers including a Saudi alignment guardrail tier; more than 100 models on HUMAIN Node. Open source

The preview has two tiers that say a good deal about the intended use. A limited preview carries what Unite.AI describes as a Saudi alignment guardrail, with thinking and streaming disabled; a research preview exposes the full checkpoint.3 Unite.AI 2026-09-03 Per benchmark scores; comparison to GPT-5.6 SOL 87.30, Opus 5 87.34, MiniMax M3 80.34; two preview tiers including a Saudi alignment guardrail tier; more than 100 models on HUMAIN Node. Open source A national model with a national alignment layer, distributed through a national API platform, is the sovereign AI product in full, and none of it required training a base model.

The benchmark claim, read carefully

The seven benchmarks span understanding, knowledge, academic exams, language proficiency, truthfulness, retrieval and a translated MMLU. The per benchmark scores range from 67.67 on Arabic EXAMS to 97.53 on AraTrust.3 Unite.AI 2026-09-03 Per benchmark scores; comparison to GPT-5.6 SOL 87.30, Opus 5 87.34, MiniMax M3 80.34; two preview tiers including a Saudi alignment guardrail tier; more than 100 models on HUMAIN Node. Open source The average lead over the two American frontier models is about two points, and the lead over the MiniMax M3 base it was built from is about nine.3 Unite.AI 2026-09-03 Per benchmark scores; comparison to GPT-5.6 SOL 87.30, Opus 5 87.34, MiniMax M3 80.34; two preview tiers including a Saudi alignment guardrail tier; more than 100 models on HUMAIN Node. Open source That second gap is the one that matters: a trillion Arabic tokens moved a strong general model nine points on Arabic tasks, which is evidence that continued pre training on a language works.

Three caveats. The evaluation is HUMAIN's, run on its own preview checkpoint, and has not been reproduced by anyone else.1 HUMAIN via PR Newswire 2026-09-03 428B mixture of experts on the MiniMax-M3 lineage; commissioned by HUMAIN, delivered by MiniMax; more than one trillion Arabic tokens; highest average on seven Arabic benchmarks; research preview on HUMAIN Node; weights targeted for next month under the MiniMax Community License; LEAP 2026 Riyadh. Open source 3 Unite.AI 2026-09-03 Per benchmark scores; comparison to GPT-5.6 SOL 87.30, Opus 5 87.34, MiniMax M3 80.34; two preview tiers including a Saudi alignment guardrail tier; more than 100 models on HUMAIN Node. Open source The two point margin over GPT-5.6 SOL and Opus 5 is within the range that prompt choices and sampling settings can produce on benchmarks of this kind. And the comparison models are general purpose systems that were not tuned for Arabic, so the result says humain-m3 is the best Arabic model on these tests, not that it is a better model.

Who gains and who loses

MiniMax gains a sovereign customer and proof that its open weight base can be the foundation of a national model, a reference no other Chinese lab can yet claim.2 Crypto Briefing 2026-09-02 23 billion active parameters; 89.37 percent average; built on the open MiniMax-M3 base; HUMAIN founded May 2025 under PIF; ALLAM 34B as prior flagship; Nvidia, AMD and Mistral partnerships; pragmatic pivot framing. Open source HUMAIN gains a frontier scale Arabic model in months rather than years, and a platform story for Node.3 Unite.AI 2026-09-03 Per benchmark scores; comparison to GPT-5.6 SOL 87.30, Opus 5 87.34, MiniMax M3 80.34; two preview tiers including a Saudi alignment guardrail tier; more than 100 models on HUMAIN Node. Open source Arabic speaking developers gain a model tuned for their language with weights on the way.1 HUMAIN via PR Newswire 2026-09-03 428B mixture of experts on the MiniMax-M3 lineage; commissioned by HUMAIN, delivered by MiniMax; more than one trillion Arabic tokens; highest average on seven Arabic benchmarks; research preview on HUMAIN Node; weights targeted for next month under the MiniMax Community License; LEAP 2026 Riyadh. Open source Riyadh gains a demonstration that its sovereign AI ambitions do not depend on either superpower's proprietary models.

The American labs lose a benchmark headline in a market they court, and the sovereign AI narrative that US export policy has tried to build around trusted American stacks loses ground: HUMAIN has partnerships with Nvidia and AMD for chips and chose China for the model.2 Crypto Briefing 2026-09-02 23 billion active parameters; 89.37 percent average; built on the open MiniMax-M3 base; HUMAIN founded May 2025 under PIF; ALLAM 34B as prior flagship; Nvidia, AMD and Mistral partnerships; pragmatic pivot framing. Open source Smaller Arabic model efforts, including HUMAIN's own ALLAM line, lose relevance. And any Gulf policymaker who imagined sovereignty meant building from scratch has a public example of the alternative.

The counter case

The assessment that this is a template could be wrong if the MiniMax dependency becomes a liability. A Chinese base under a Chinese community license, further trained by a Chinese lab, sits awkwardly with a Saudi compute strategy built on American chips, and US policy toward Chinese models used in allied national infrastructure is not settled.2 Crypto Briefing 2026-09-02 23 billion active parameters; 89.37 percent average; built on the open MiniMax-M3 base; HUMAIN founded May 2025 under PIF; ALLAM 34B as prior flagship; Nvidia, AMD and Mistral partnerships; pragmatic pivot framing. Open source If Washington attaches conditions to chip access that touch model provenance, HUMAIN's shortcut becomes expensive. The benchmark lead may also not survive independent evaluation, in which case the story is a strong Arabic fine tune rather than a frontier result. And the weights release is a target, not a fact; a sovereign model that stays behind a national API is a different thing from one the region can build on.1 HUMAIN via PR Newswire 2026-09-03 428B mixture of experts on the MiniMax-M3 lineage; commissioned by HUMAIN, delivered by MiniMax; more than one trillion Arabic tokens; highest average on seven Arabic benchmarks; research preview on HUMAIN Node; weights targeted for next month under the MiniMax Community License; LEAP 2026 Riyadh. Open source

What to watch

  • Weights actually ship. A public release under the MiniMax Community License by mid October would confirm the open commitment; a slip past the end of 2026 would suggest the guardrail tier is the real product.1 HUMAIN via PR Newswire 2026-09-03 428B mixture of experts on the MiniMax-M3 lineage; commissioned by HUMAIN, delivered by MiniMax; more than one trillion Arabic tokens; highest average on seven Arabic benchmarks; research preview on HUMAIN Node; weights targeted for next month under the MiniMax Community License; LEAP 2026 Riyadh. Open source
  • Independent Arabic evaluation. A third party reproducing the 89.37 average, or a materially different figure, within three months settles whether the lead is real.3 Unite.AI 2026-09-03 Per benchmark scores; comparison to GPT-5.6 SOL 87.30, Opus 5 87.34, MiniMax M3 80.34; two preview tiers including a Saudi alignment guardrail tier; more than 100 models on HUMAIN Node. Open source
  • A second sovereign model on a Chinese base. Another Gulf, Asian or African national model announced on MiniMax, Qwen, DeepSeek or Kimi weights within a year would confirm the template.2 Crypto Briefing 2026-09-02 23 billion active parameters; 89.37 percent average; built on the open MiniMax-M3 base; HUMAIN founded May 2025 under PIF; ALLAM 34B as prior flagship; Nvidia, AMD and Mistral partnerships; pragmatic pivot framing. Open source
  • US policy on model provenance. Any export or partnership condition tying American chip access to the origin of models run on them, within a year, would turn HUMAIN's choice into a test case.2 Crypto Briefing 2026-09-02 23 billion active parameters; 89.37 percent average; built on the open MiniMax-M3 base; HUMAIN founded May 2025 under PIF; ALLAM 34B as prior flagship; Nvidia, AMD and Mistral partnerships; pragmatic pivot framing. Open source

Sovereign AI, in this instance, means Saudi data, Saudi alignment and Saudi distribution on a Chinese foundation. The region will notice how well that works before Washington decides how it feels about it.