Saudi Arabia’s National Ai Model Runs on Chinese Weights

At LEAP in Riyadh on 3 September, Saudi Arabia’s state-backed AI company HUMAIN launched humain-m3 as the backbone of a national AI platform: Arabic-first, built for local language and culture, and the technical centrepiece of the kingdom’s sovereignty ambitions.

It is a 428-billion-parameter mixture-of-experts model, meaning it carries an enormous number of parameters but routes each request through only about 23 billion of them at a time. That is what keeps a model of this size affordable to serve.

Those figures match MiniMax M3, which the Shanghai company released as open weights on 1 June, and so does the architecture. HUMAIN trained that base on more than a trillion further Arabic tokens.

One detail in the announcement matters more than the parameter count. According to Bloomberg’s report and Al-Monitor’s account of the launch, HUMAIN said it had commissioned MiniMax to deliver the model. This was not simply a matter of downloading published weights and working on them independently. It was a commercial engagement with the Chinese laboratory.

HUMAIN has not hidden the relationship. It describes the model as open weight, its leadership has argued publicly that open weights are the route to sovereignty, and the MiniMax connection was named in the launch reporting.

So the interesting question is not whether the choice was disclosed. It is what the choice tells you about where AI capability is actually being sourced — and what “sovereign” means when the supplier is a foreign company you have contracted with.

Why start from someone else’s model

Training a frontier model from scratch takes years, an enormous compute budget and a research team that does not exist in most countries. Post-training an existing one on a trillion tokens of your own language takes months.

HUMAIN had already been down the first road. Its previous flagship was ALLAM 34B, a considerably smaller Arabic system. Moving to a 428-billion-parameter base it did not build represents a decision that speed to a competitive model matters more, for now, than owning the architecture.

The strategic argument its leadership has made is that open weights are insurance. A hosted model from an American or Chinese provider can be withdrawn, restricted by export policy, or priced differently tomorrow. Weights sitting on your own hardware in your own data centre cannot be switched off remotely. On that reading, downloading a capable open model is not a compromise on sovereignty — it is the mechanism that delivers it.

The context around the launch supports the seriousness of the ambition. Saudi Arabia announced more than $15 billion in AI infrastructure agreements on the opening day of LEAP 2026, including a $10 billion arrangement involving AMD and Cisco and a $5 billion AWS zone. HUMAIN, owned by the Public Investment Fund, sits at the centre of nearly all of it.

The assumption underneath the strategy

The insurance argument holds only while capable models keep being released under terms that permit this kind of use. That condition is no longer something a planner can take for granted.

Release	Date	Licence position
MiniMax M2	October 2025	MIT licence
MiniMax M2.5	February 2026	MIT licence
MiniMax M2.7	March 2026	Modified-MIT applied after the weights were published: non-commercial use free, commercial use requires prior written authorisation and a “Built with MiniMax M2.7” attribution
MiniMax M3	June 2026	Released as open weights — the base for humain-m3

Read in order, that table does not show a company steadily closing down. MiniMax restricted M2.7 in the spring, then released M3 as open weights in June. What it shows is something more awkward for anyone planning around it: licensing has become case-by-case rather than a standing policy.

MiniMax built its developer reputation on permissive releases. Then, months after listing on the Hong Kong Stock Exchange in January 2026, it revised the terms on one of them after the weights had already gone up on Hugging Face. Research, personal projects and private fine-tuning were unaffected. Hosted services and commercial products moved into territory needing the company’s written permission, plus a visible “Built with MiniMax M2.7” credit.

MiniMax’s stated reason was quality control: preventing intermediaries from serving degraded versions of the model to developers under its name. Whether or not that justifies the mechanism, it is a different motive from simply closing a model down, and the company said the authorisation process would be quick.

Nor is it only MiniMax. The same discussion pointed to comparable moves on Qwen 3.6 and GLM-5.1. Alibaba’s Qwen team is reported to have shifted toward proprietary development following senior departures, and Xiaomi released its MiMo v2 models under a closed licence. The familiar shorthand — that Chinese laboratories publish weights while American ones do not — no longer describes the landscape reliably.

The scores have not been independently checked

HUMAIN reports an average of 89.37% across seven public Arabic benchmarks, describing it as the strongest result among the frontier models it tested. The benchmarks are established ones and the post-training work behind the figure appears substantial.

The model is available initially as a research preview through HUMAIN Node, the company’s platform for giving developers, researchers and enterprises access to models and inference. Alongside it HUMAIN introduced a conversational product covering Saudi, Maghrebi, Egyptian and Levantine Arabic dialects.

The evaluations were run on HUMAIN’s own infrastructure. They have not been submitted to the Open Arabic LLM Leaderboard, which is where comparable regional systems post results — the UAE’s Falcon-H1 Arabic and Qatar’s Fanar among them. Until that happens there is no like-for-like comparison with the models humain-m3 is implicitly measured against.

Self-reported benchmarks are normal at launch and are not evidence of anything improper. They are simply not the same as a ranking.

The geopolitical knot

One further complication sits underneath the technical story.

In April 2026 the House Homeland Security Committee and the Select Committee on China opened a joint investigation into national security risks from Chinese AI models, naming MiniMax alongside DeepSeek, Alibaba and Moonshot AI. Saudi Arabia is meanwhile a significant customer for American AI hardware, and HUMAIN’s infrastructure deals run through American vendors.

The relationship is not new. MiniMax’s president and co-founder met Saudi Arabia’s communications and information technology minister in July to discuss cooperation across generative AI. Separately, HUMAIN and the American infrastructure company Together AI announced a 250 MW Saudi data centre aimed specifically at serving open models.

So the kingdom is building national AI capability on American silicon and Chinese model architecture simultaneously, while the laboratory supplying that architecture is under scrutiny in Washington. Whether that becomes a practical problem or stays an awkward footnote is not something anyone can call yet.

The pattern worth watching

Strip away the specifics and a structural shift is visible. Countries that want their own AI capability, and cannot spend years building a frontier model, are reaching for whatever capable open model exists — and increasingly that model is Chinese rather than American.

Saudi Arabia is not the only country working through this. Russia’s largest bank built GigaChat under different constraints and reached a comparable conclusion, benchmarking its flagship against a Chinese open-weight model rather than an American one. The pattern is the same wherever a national project needs a capable base it did not build.

Chinese laboratories have moved at pace. As the European Commission’s own analysis of the open-weight race sets out, Z.ai has published GLM-5.2 under an MIT licence, Moonshot’s Kimi K3 runs to 2.8 trillion parameters, and Alibaba released Qwen3.8 at 2.4 trillion. Whatever the merits of any individual model, that release cadence is what makes a Chinese base the obvious pragmatic choice for a national project in a hurry.

The open question is how durable that foundation is. HUMAIN’s bet is that open weights buy independence, and for the model it already holds, the bet is sound. Whether it holds for the next one rests with companies that are publicly listed, commercially pressured, under political scrutiny in Washington, and that have already licensed one model openly and restricted another within the same season.

Frequently asked questions