Start with the memory arithmetic, because it explains who got to be the local partner. HUMAIN says the mixture-of-experts routing activates 23 billion parameters per token [4], roughly 5.4 percent of the parameters in the checkpoint [16]. Compute per token is sized like a 23B model while the resident weights are sized like a 428B one. HUMAIN was launched by Saudi Arabia's Public Investment Fund with data centers and cloud infrastructure written into its remit [6], and it offers global, in-Kingdom and sovereign hosting as deployment options [8]. That infrastructure is the real currency in this arrangement, and no fine-tuning budget covers it.
The Arabic layer sits on machinery HUMAIN did not design. MiniMax describes M3 as carrying a context window of up to 1 million tokens, native image and video inputs, computer use and an architecture it calls MiniMax Sparse Attention, and it positioned the model around coding and long-running agents [9]. HUMAIN M3 and MiniMax M3 differ by one word in the name and an Arabic corpus in the tail [2]. Long-context behaviour, attention sparsity and tool-use conventions arrive as inherited defaults, fixed well before post-training on Arabic tokens ever begins.
Trying it is cheap. The endpoint is OpenAI-compatible and the model string is humain-m3 [3], so the client change is a base URL and a name. The evaluation is the expensive part, and the preview structure complicates it: HUMAIN runs a guardrailed checkpoint for everyday testing alongside a research preview with thinking, streaming and lower latency [5]. A latency or quality figure taken from one checkpoint describes only that checkpoint, not the other. The supplied announcement carries no independent benchmark table or evaluation methodology, so the performance claims remain company-reported [10]. For any of them to transfer, they would have to be measured on the dialects in your own traffic, on the checkpoint you would actually be served, with agent tasks that look like yours. MiniMax's own cost demonstrations have the same shape: RuntimeWire reported in August that MiniMax said M3 finished a business-email task for $0.018, while noting the visible demonstration did not establish a repeatable cost [11].
For MiniMax, licensing a base is the channel that requires the least local work. Its prospectus reported $53.4M in revenue for the nine months ended September 30th, 2025, against $30.5M for full-year 2024 [12]; annualising the nine-month figure gives about $71M, roughly 2.3 times last year [17]. It reports more than 100,000 registered enterprise customers and developers across more than 100 countries and regions [13], from a company Junjie Yan founded in 2021 after more than six years at SenseTime [14]. Selling the foundation to a state-backed partner buys distribution in a market where data residency and dialect quality decide procurement, and MiniMax picks up that distribution without doing any Arabic data collection of its own.
HUMAIN M3 is roughly 12.6 times the parameter count of the ALLAM 34B product HUMAIN shipped in August [18], reached without starting a foundation-model training run of its own [20]. What the local partner owns is the Arabic corpus, the alignment and dialect testing, the hosting boundary and the approval list [8][19].
