MiniMax M3 powers HUMAIN's 428B-parameter Arabic model

Why it matters

HUMAIN M3 shows how frontier labs can export foundation models without owning the local product: MiniMax supplies the base, while HUMAIN controls Arabic training, access and hosting.

MiniMax founder Junjie Yan has found a new route into the Middle East: Saudi Arabia's HUMAIN used MiniMax M3 as the foundation for HUMAIN M3, a 428-billion-parameter model post-trained on more than 1 trillion Arabic tokens.

MiniMax said in a post on X on September 3rd that M3 powers the Arabic-focused model. HUMAIN M3 is live in limited preview through a playground and an OpenAI-compatible API under the model name humain-m3. HUMAIN says the mixture-of-experts architecture activates 23 billion parameters for each token and accepts text, images and video.

HUMAIN on X

The release turns Yan's general-purpose model into the base layer for a sovereign Arabic system. MiniMax supplies the expensive foundation-model work. HUMAIN adds Arabic data, dialect testing, regional alignment and deployment through infrastructure that offers global, in-Kingdom and sovereign hosting options.

That division of labor gives both sides something they need. HUMAIN can field a much larger model without beginning another foundation-model training run from zero. MiniMax gains a state-backed distribution partner in a market where language quality, cultural context and data residency can matter as much as raw benchmark scores.

Yan's model becomes an export layer

Yan founded MiniMax in 2021 after spending more than six years at SenseTime, where he served as a vice president and deputy head of the research institute. A mathematician by training, he earned a doctorate in artificial intelligence from the Chinese Academy of Sciences' Institute of Automation and later conducted postdoctoral research at Tsinghua University.

His work at MiniMax has paired model research with aggressive commercialization. MiniMax's prospectus reported $53.4M in revenue for the nine months ended September 30, 2025, compared with $30.5M for full-year 2024. MiniMax also reported more than 100,000 registered enterprise customers and developers across more than 100 countries and regions.

MiniMax describes M3 as having a context window of up to 1 million tokens, native image and video inputs, computer use and an architecture called MiniMax Sparse Attention. MiniMax positioned M3 around coding and long-running agents, though HUMAIN is using the same base machinery for a language and regional specialization project.

Yan has spent recent weeks putting practical examples around M3's efficiency claims. RuntimeWire reported in August that MiniMax said M3 completed a business-email task for $0.018, while noting that the visible demonstration did not establish a repeatable cost. HUMAIN M3 is a more consequential test of the same strategy: another organization has taken Yan's model, trained it further and attached its own flagship name.

That is a useful commercial role for a frontier lab outside the largest US cloud platforms. A base model can be sold directly through tokens, embedded inside developer products, or licensed as the starting point for national and industry-specific systems. HUMAIN M3 puts MiniMax in the third category.

HUMAIN supplies the Arabic layer

The supplied announcement does not provide an independent benchmark table or evaluation methodology, so HUMAIN M3's performance claims remain company-reported. Preview users will provide the first external test of its dialect coverage, Arabic-English interaction and agent capabilities.

Preview access requires approval. HUMAIN offers a guardrailed checkpoint for everyday testing and a separate research preview with thinking, streaming and lower latency. HUMAIN says feedback from those users will shape later checkpoints before the weights are opened.

HUMAIN entered the market with a national mandate. Saudi Arabia's Public Investment Fund launched HUMAIN on May 12th, 2025 to operate across data centers, cloud infrastructure, models and applications. HUMAIN had already inherited the ALLAM Arabic model line and released an ALLAM 34B-powered chat product in August 2025.

HUMAIN CEO Tareq Amin brings a telecom operator's approach to that mandate. Before joining HUMAIN in May 2025, Amin led Aramco Digital, launched Rakuten Symphony and ran Rakuten Mobile. His career has centered on deploying large technical systems through partnerships, which helps explain the structure of the MiniMax arrangement: HUMAIN controls the product, Arabic post-training and deployment layer while importing a foundation model from a specialist lab.

Arabic AI has become a regional contest

HUMAIN M3 enters a field already crowded with government-backed Arabic models. In December 2025, Inception, Cerebras and Mohamed bin Zayed University of Artificial Intelligence released Jais 2, a 70-billion-parameter open-weight model trained for Modern Standard Arabic, regional dialects and code-switching.

Abu Dhabi's Technology Innovation Institute followed in January 2026 with Falcon-H1 Arabic, a family of 3B, 7B and 34B models using a hybrid Mamba-Transformer architecture. TII emphasized smaller deployment footprints and context windows reaching 256,000 tokens.

HUMAIN is choosing scale and multimodality. Its 428B mixture-of-experts design is far larger in total parameters than Jais 2 or Falcon-H1 Arabic, while the MiniMax foundation gives HUMAIN M3 coding, tool use, computer control and image and video understanding. Parameter counts do not establish quality, and the models target different deployment budgets. HUMAIN's bet is that a large imported foundation plus intensive Arabic post-training can move faster than building every layer domestically.

The result also clarifies MiniMax's international strategy under Yan. MiniMax does not need its own consumer brand to win every market. It can become the model underneath institutions that already possess distribution, infrastructure and local data. HUMAIN M3 is the clearest version of that strategy so far: Saudi ownership and Arabic training on top, with Yan's Shanghai-built model doing the foundational work underneath.