Uncategorized

Saudi Arabia’s HUMAIN launches humain-m3, an Arabic AI model built with China’s MiniMax

0
Please log in or register to do it.
HUMAIN Chat

HUMAIN, the AI company owned by Saudi Arabia’s Public Investment Fund, announced humain-m3 on 3 September 2026: a 428-billion-parameter Arabic-language model commissioned by HUMAIN and developed by the Chinese AI company MiniMax. It is available in research preview on the HUMAIN Node platform, and HUMAIN says open weights will follow once safety training is complete.

Last updated: September 2026

What HUMAIN announced

HUMAIN unveiled the model at the LEAP technology conference in Riyadh. According to the company’s press release, humain-m3 is a mixture-of-experts model built on MiniMax-M3 and pre-trained on more than one trillion tokens of Arabic-native content. HUMAIN also develops the ALLaM family of Arabic models.

HUMAIN chief executive Tareq Amin said Arabic remains significantly underrepresented at the frontier of AI despite hundreds of millions of speakers. The preview lets researchers and developers test the model and give feedback before the open-weight release.

humain-m3 specs: what the model can do

HUMAIN Node and MiniMax’s model card for MiniMax-M3 give the technical picture:

  • Size: 428 billion total parameters, with about 23 billion active per token.
  • Modalities: text, image and video, trained jointly from the start, according to HUMAIN Node.
  • Thinking modes: three settings, always-on, adaptive and off, so developers can trade reasoning depth for speed.
  • Context window: MiniMax lists 1 million tokens for the base MiniMax-M3 model; HUMAIN has not published a separate figure for humain-m3.
  • Developer access: a playground and an API with OpenAI-compatible SDKs, using the model name humain-m3.

How humain-m3 scores on Arabic benchmarks

HUMAIN reports that humain-m3 had the highest average score among the frontier models it tested across seven public Arabic benchmarks. The figures below are HUMAIN’s own, published on HUMAIN Node, and compare against models it labels GPT-5.6 SOL and Opus 5.

Benchmarkhumain-m3GPT-5.6 SOLOpus 5MiniMax-M3 reference
AlGhafa86.45%81.54%83.06%75.31%
ArabicMMLU90.70%88.82%88.77%81.80%
Arabic EXAMS67.67%66.40%64.80%61.27%
MadinahQA95.44%94.48%94.78%87.10%
AraTrust97.53%93.42%91.38%90.60%
ALRAGE94.63%94.91%94.85%79.32%
Translated MMLU93.20%91.48%93.78%87.00%
Average89.37%87.30%87.34%80.34%

Tech Times noted that the evaluations ran on HUMAIN’s own infrastructure and were not submitted to the independent Open Arabic LLM Leaderboard. Until outside testing is published, the lead should be read as a vendor claim.

Why it matters for MENA users and businesses

For Arabic-speaking teams, the practical point is a large Arabic-focused model that may become downloadable. If the weights ship, companies could run it on their own servers or a regional cloud instead of a US API, which matters for sectors with data-residency rules such as government, banking and healthcare.

The deal also shows Gulf AI programmes mixing suppliers: HUMAIN has large infrastructure partnerships with US firms, including Nvidia and AWS, which it expanded in November 2025, yet chose a Chinese lab for this model. Al-Monitor described the launch as putting Saudi Arabia at the center of the technology race between Washington and Beijing. Tech Times reported that the US House Homeland Security Committee and the House Select Committee on China opened a joint investigation into MiniMax in April 2026, and that HUMAIN Node runs on hardware located in Saudi Arabia, so prompts are not routed to MiniMax’s servers in China. Dimsum Daily reported that MiniMax’s Hong Kong-listed shares rose about 10% on 4 September, the day after the announcement.

Regional alternatives exist: Abu Dhabi’s Technology Innovation Institute develops the Falcon models, including Falcon-H1 Arabic, launched in January 2026. General-purpose assistants such as ChatGPT and Claude also remain options for Arabic-speaking users.

What happens next

HUMAIN said the weights would be released under the MiniMax Community License once safety training and alignment are finished, targeting October 2026. At the time of writing, HUMAIN Node still lists open weights as coming soon. Tech Times reported that revenue-generating deployments will need a separate commercial agreement. HUMAIN has not published pricing for the preview.

Frequently asked questions

What is humain-m3?

humain-m3 is an Arabic-focused large language model from Saudi Arabia’s HUMAIN, developed by the Chinese company MiniMax. It is a 428-billion-parameter mixture-of-experts model built on MiniMax-M3 and pre-trained on more than one trillion tokens of Arabic content, announced on 3 September 2026.

How can I try humain-m3?

Create an account on HUMAIN Node at node.humain.com; according to Unite.AI, access is granted after approval. The limited preview adds Saudi alignment guardrails and disables thinking and streaming; the research preview offers the full checkpoint with thinking and streaming enabled, through a playground and an OpenAI-compatible API.

Is humain-m3 open source?

Not yet. HUMAIN plans to publish the weights under the MiniMax Community License, targeted for October 2026, after safety training. That licence is not a standard open-source licence, and Tech Times reported that commercial, revenue-generating use will need a separate agreement.

Is humain-m3 better than ChatGPT or Claude in Arabic?

HUMAIN reports an 89.37% average across seven Arabic benchmarks, ahead of models it labels GPT-5.6 SOL (87.30%) and Opus 5 (87.34%). These are self-reported results, not independently verified, and humain-m3 did not lead on two of the seven tests, ALRAGE and Translated MMLU.

Does humain-m3 send data to China?

According to Tech Times, HUMAIN Node runs on hardware in Saudi Arabia, so prompts are not routed through MiniMax’s servers in China. Businesses with strict compliance needs should still review HUMAIN’s terms, and can consider self-hosting once weights are released.

Sources

ChatGPT vs Claude vs Gemini (September 2026): Which AI Assistant Is Best?
OpenAI Launches GPT-6 Sol and Luna, Cutting API Prices in Half
Ad Area

Reactions

0
0
0
0
0
0
Already reacted for this post.

Reactions