Saudi Arabia's HUMAIN ships its first Arabic LLM — on top of China's MiniMax
Saudi Arabia's state-backed AI company HUMAIN released the first large language model purpose-built for Arabic, and the foundation underneath isn't Saudi at all: it's MiniMax's open-weights M3 stack. The Riyadh-based unit of the Public Investment Fund disclosed the model on its developer platform Humain Node, framing it as the backbone of a national AI platform under the kingdom's Vision 2030 diversification push. The pivot matters because HUMAIN was previously promoting ALLAM 34B as its flagship Arabic system; rebuilding on a Chinese open-weights line is the clearest signal yet that sovereign-AI buyers are choosing speed-to-market over from-scratch development.
HUMAIN M3 is a 428-billion-parameter mixture-of-experts model with 23 billion parameters active per token. HUMAIN further pre-trained it on more than one trillion tokens of Arabic-native content, and the company says it averaged 89.37 percent across seven Arabic benchmark tests — claiming a win over GPT-5.6 SOL on that suite. MiniMax, the Shanghai-based AI lab that IPO'd in Hong Kong in January at HKD165 and closed today at HKD361.40, has now landed a sovereign-AI customer in a Gulf monarchy. HUMAIN says it plans to open-source the resulting weights and deploy them across Saudi sovereign infrastructure; Humain Node gives researchers early hands-on access today.
The geopolitics are doing most of the work. Saudi Arabia already runs partnerships with Nvidia, AMD and Mistral; picking a Chinese open-weights base for the kingdom's flagship Arabic model is the kind of choice that gets read in Washington. The defensible framing is the one HUMAIN and MiniMax are using: open weights aren't a proprietary dependency, and the heavy lift — Arabic data curation, evaluation, deployment — is genuinely Saudi. That argument lands in Beijing and Riyadh. It lands less cleanly in Brussels and the US, where the export-control conversation has been quietly accelerating all year.
Microsoft's Project Zenith is a 'distraction-free' Windows for local 30B+ AI models
Microsoft previewed Project Zenith, a developer-focused Windows experience aimed at running large AI models — 30 billion parameters and up — locally on machines with 64GB of memory or more. Tom Warren of The Verge reports the pitch is a stripped-down shell that pulls the browser, notifications and other ambient noise out of the way so the model, and the developer's interaction with it, can have the machine to themselves. It arrives the same week Nvidia and Lenovo showed a 120-billion-parameter model running on a laptop with RTX Spark, and the timing isn't accidental: the local-AI story is now an OS story as much as a hardware one.
Microsoft's bet is that the developer audience — the people writing agents, fine-tunes, and tool-calling pipelines — will trade the familiar Windows desktop for a calmer environment if it means a 30B-class model is one click away with no rate limits, no telemetry, and no cloud bill. The 64GB memory floor is generous but not exotic in 2026; Lenovo's ThinkPad P-series, Apple's M4 Max workstations, and most gaming-tier NUCs already clear it. The bigger constraint will be thermals: a 30B-parameter inference at usable token rates is a sustained, heavy workload, and Zenith is implicitly betting that silicon and chassis design have caught up with what local AI wants to do.
The competitive read is sharper than the product read. OpenAI, Anthropic and Google are all leaning hard into the agent-and-cloud model where the context, the tools, and the long-running sessions live on their side of the wire. Microsoft, post its deepening OpenAI relationship, is now also building the path that lets a developer run a frontier-sized model fully outside that loop. Project Zenith is the company's answer to the question: what does Windows become in a world where the most important thing it runs is an open-weights model, not Office?
Are sovereign-AI buyers going to keep defaulting to Chinese open-weights bases, or will the geopolitical pressure finally outweigh the speed-to-market advantage? Tell us in the comments.
