Saudi Arabia's New Arabic Model Runs on China's MiniMax

Saudi Arabia has a new flagship Arabic-language AI model, and the most telling thing about it is whose technology sits underneath. On September 3, at the LEAP conference in Riyadh, the state-backed company HUMAIN unveiled humain-m3, a frontier model it built on the lineage of MiniMax, a Chinese lab, rather than on any American system.

The specifications are serious. humain-m3 is a mixture-of-experts model with roughly 428 billion total parameters and about 23 billion active for any given token, a design that keeps running costs down while holding a large store of knowledge. HUMAIN took the MiniMax base and continued training it on more than a trillion tokens of Arabic-native text. On seven public Arabic benchmarks the company says the model posted the highest average of any frontier system it tested, around 89 percent.

A pragmatic choice with a geopolitical edge

HUMAIN is backed by the Public Investment Fund, the sovereign wealth vehicle at the centre of Saudi Arabia's push to become an AI power in its own right. Building on a Chinese model is a deliberate call. As Bloomberg reported, the decision signals a pragmatic streak in how the kingdom is approaching sovereign AI: take the strongest available open base, wherever it comes from, and adapt it rather than wait for a home-grown one. MiniMax's Hong Kong-listed shares climbed on the news.

That choice lands in a charged context. Washington has spent the past year tightening the flow of advanced AI technology to some partners while courting others, and Gulf states have become a prize in that contest. A high-profile Saudi model built on Chinese rather than American foundations is a quiet reminder that when the United States makes its systems harder to obtain, capable alternatives are a download away. It is the same tension now shaping the debate over chip tariffs and export rules.

How it will be used

humain-m3 is available in research preview through HUMAIN Node, the company's cloud platform, in two tiers. A limited-preview version ships with a Saudi alignment guardrail applied to its outputs. A research-preview version offers the full checkpoint with reasoning traces, streaming, and lower latency. Positioning the model as an Arabic-first system is its own strategy. The largest Western and Chinese models treat Arabic as one language among hundreds, while humain-m3 is tuned to serve roughly half a billion Arabic speakers as a primary audience. It joins a widening field of capable, cheaper, more specialised models arriving almost weekly, from Microsoft's bargain speech engine to a steady run of open releases. The interesting question now is whether a nationally aligned model, trained for one language and one set of values, becomes a template other governments copy.