The press release said the quiet part in its own headline. "HUMAIN Unveils humain-m3, a Frontier Arabic Language Model Developed by MiniMax" — that is the actual title of the announcement the Public Investment Fund's artificial intelligence company put on the wire from Riyadh on Wednesday, September 3, during LEAP 2026. MiniMax is a Shanghai company listed in Hong Kong under stock code 00100. The flagship Arabic model of a kingdom that has spent two years lobbying Washington for Nvidia allocations was commissioned from a Chinese lab.
The numbers are specific. humain-m3 is a 428-billion-parameter mixture-of-experts model built on the MiniMax-M3 lineage, activating roughly 23 billion parameters per token, and further pre-trained on more than one trillion tokens of Arabic-native content. It is live now as a research and evaluation preview on HUMAIN Node, the company's model-access platform, reachable through a no-code playground or an OpenAI-compatible API endpoint. HUMAIN says it expects to release the weights under the MiniMax Community License once safety training and alignment are finished, currently targeted for next month.
MiniMax shareholders liked it. Dimsum Daily reported the Hong Kong-listed stock jumped about 10 percent, vindicating a strategy that has nothing to do with selling API calls: supply the base model, let sovereign operators bolt their own language data on top.
What HUMAIN actually shipped
The base architecture is not new. MiniMax released M3 as open weights on June 1, 2026 — a one-million-token context window, native text, image and video input, and a sparse attention mechanism that indexes 2,048-token blocks rather than running full attention across the whole window. HUMAIN commissioned MiniMax to deliver the Arabic-tuned variant, then continued pre-training it.
The Arabic layer is HUMAIN's contribution, and it is not trivial. Curating that corpus, running continued pre-training at frontier scale, and building dialect coverage — HUMAIN Voice spans Saudi, Maghrebi, Egyptian and Levantine Arabic, four families that Modern Standard Arabic-trained systems handle badly — is real work. Access is tiered: a limited preview applies a Saudi alignment guardrail with thinking and streaming disabled, while a research-preview tier serves the full checkpoint.
HUMAIN's previous flagship, ALLAM, is a 34-billion-parameter model. Going from 34B to 428B from scratch is a several-hundred-million-dollar, multi-year problem. Renting the architecture solved it in months.
The benchmark claim is HUMAIN's own
HUMAIN says humain-m3 posted the highest average score among the frontier systems it tested across seven public Arabic benchmarks. Read that sentence carefully: it tested them, on its own infrastructure.
The published figures put humain-m3 at an 89.37% equally weighted average — 86.45% on AlGhafa, 90.70% on ArabicMMLU, 67.67% on Arabic EXAMS, 95.44% on MadinahQA, 97.53% on AraTrust, 94.63% on ALRAGE and 93.20% on Translated MMLU. In the same run, GPT-5.6 SOL scored 87.30%, Opus 5 scored 87.34%, and the MiniMax M3 reference checkpoint scored 80.34%. HUMAIN says its Arabic post-training adds about nine percentage points over that reference.
The seven tasks are the same suite behind the Open Arabic LLM Leaderboard, the Hugging Face-maintained pipeline where Arabic models are independently scored. As of publication, HUMAIN has not submitted humain-m3 to it. That matters because the regional comparison points were all scored there: the Technology Innovation Institute's Falcon-H1 Arabic 34B at 75.36%, Qatar Computing Research Institute's Fanar-2-27B-Instruct at 69.40%, and the academic Arabic-DeepSeek-R1 at 80.18%. Two evaluation environments, no shared baseline. Tech Times reporter Joshua Mitchell, who first flagged the gap, noted that even backend differences — vLLM versus the leaderboard's Accelerate implementation — can move Arabic benchmark scores by roughly a point, which is meaningful at these margins.
What "sovereign" means when the base is foreign
"Arabic is spoken by hundreds of millions of people, yet it remains significantly underrepresented at the frontier of artificial intelligence," said Tareq Amin, HUMAIN's CEO, in the announcement. "With humain-m3, we are investing in changing that. And through HUMAIN Node, we are making that intelligence accessible so developers, researchers and innovators can experiment with it, build on it and create the next generation of Arabic AI experiences."
The representation gap is real. The strategic read is that closed American frontier labs cannot serve this customer. OpenAI and Anthropic sell access, not weights; a state-backed company cannot download a GPT checkpoint, host it in Riyadh, retrain it and call it national infrastructure. Chinese labs — MiniMax, DeepSeek, Alibaba's Qwen, Moonshot — will let you do exactly that. Startup Fortune's Dave Barr put the pattern bluntly: sovereign AI will be built with "American silicon where possible, local data where available — and open Chinese weights whenever they get the job done."
That sits awkwardly against the chip file. In November 2025 the Commerce Department authorized HUMAIN to buy up to 35,000 Nvidia GB300-equivalent chips on a case-by-case basis with security and reporting conditions. On July 10, 2026, Commerce elevated the UAE to Country Group A:5, letting G42 and Core42 receive advanced chips without a license. Saudi Arabia got no equivalent rule. Meanwhile, in April 2026 the House Homeland Security Committee and the Select Committee on China opened a joint investigation into PRC-developed AI models, naming MiniMax alongside DeepSeek, Alibaba and Moonshot.
One nuance most coverage skipped: HUMAIN Node runs on Saudi hardware — roughly 18,000 Nvidia GB300s, a 13,000-GPU AMD inference cluster due by Q4, and 19,000 Groq LPUs — so prompts do not transit MiniMax's servers. The dependency is architectural and contractual, not a data pipe. The MiniMax Community License still requires a separate commercial agreement for revenue-generating deployments, meaning Saudi Arabia's national Arabic model carries a Chinese licensing counterparty.
What to watch
The open-weight release, targeted for October, is the resolution point: once the weights are public, anyone can run humain-m3 through the OALL pipeline and check the 89.37% against Falcon-H1 Arabic and Fanar on identical infrastructure. Also worth tracking: whether the commercial terms in the MiniMax license constrain Gulf enterprise deployment; whether Washington's MiniMax investigation touches allied procurement; and whether Abu Dhabi's TII, which trains Falcon from scratch, treats Riyadh's shortcut as a rebuke or a template.
“Arabic is spoken by hundreds of millions of people, yet it remains significantly underrepresented at the frontier of artificial intelligence.”— Tareq Amin, CEO, HUMAIN