Picture a Riyadh evening in mid-November. Inside a tower office, the chief executive of Saudi Arabia’s national AI company steps up to a microphone and makes a promise so large it sounds almost fictional. By December 15, he says, his country will have something nobody else on the planet has.
The promise came from Tareq Amin, chief executive of HUMAIN. His company, owned by Saudi Arabia’s sovereign wealth fund, says its full AI stack will be live in the Kingdom on December 15, centred on a 3 trillion parameter open-weight AI model that Amin calls the largest of its kind in the world.
If the model arrives as described, it would be a genuine milestone. It would also be a test of how much the world should believe before it sees the evidence.
What HUMAIN Announced in Riyadh
Amin made the announcement on Wednesday evening at HUMAIN’s offices in Majdoul Tower, Riyadh, according to Arab News. The event also unveiled HUMAIN+, an applied AI education and training platform that is due to begin operations in the first quarter of 2027.
The December date is not a vague target. Amin gave it in passing, almost casually: “The complete stack, December 15, is not far away.” The stack, as he described it, covers everything from the model itself to the cloud inference layer and the physical infrastructure underneath it.
That is the core of the pitch. Saudi Arabia is not just buying a model. It is building a sovereign AI stack that is designed, hosted and operated inside the country, so government and business users never need to send their data abroad.
“The software stack is built by Saudis, the cloud inferencing is built by Saudis, infrastructure deployed, operated by Saudis, nobody has done this,” Amin said at the event.
A 3 Trillion Parameter Model With No Name Yet
The headline number is the parameter count. Three trillion parameters would make HUMAIN’s model the largest open-weight model ever announced, if the weights are actually released and the count is real.
For comparison, Meituan’s LongCat-2.5, released under the MIT licence in November, carries 1.6 trillion total parameters. Kimi’s K3.1 Pro sits near 2.8 trillion. A 3 trillion parameter model would edge past both, though size alone says little about quality.
Alibaba’s Qwen4-Max, meanwhile, remains in a short preview window with no open weights at all. HUMAIN’s promise, if kept, would be a different kind of release entirely.
Here is the catch. The model has no name. There is no model card, no benchmark score, no published architecture and no pricing. Arab News reported the announcement without independently checking any of the claims, and no independent benchmark has been published.
That does not make the claim false. It makes it unverified. Until December 15, the world’s largest open-weight model exists only as a promise from a stage in Riyadh.
The 88% Claim Nobody Can Check
Amin’s boldest number is not the parameter count. He said the model uses 88 percent fewer tokens and costs 88 percent less than models available outside Saudi Arabia.
He did not say which models he was comparing it with. He did not explain how the saving is measured, or whether it applies to training, inference, or both. Without a baseline, the figure is a marketing line, not a measurement.
This is a familiar pattern in AI announcements. A striking percentage is easy to remember and hard to disprove until the product exists. The honest test will come when developers can actually run the model and compare it themselves.
That test is exactly what December 15 is for. If HUMAIN publishes the weights and the performance figures, the 88 percent claim can be checked in days. If it does not, the claim will hang in the air.
Why Saudi Arabia Wants Its Own AI Stack
To understand the announcement, it helps to know what HUMAIN is. The company was set up in May 2025 by Saudi Arabia’s Public Investment Fund to cover data centres, cloud platforms and AI models.
Its stated ambition is large. Amin wants Saudi Arabia to become the world’s third-largest AI player, after the United States and China. The company has announced partnerships with Nvidia, AMD, Cisco, AWS, Microsoft and xAI, and it expects 250 megawatts of data centre power by early 2027, rising to six gigawatts by 2034.
The sovereign angle matters. For governments and large enterprises, running AI on infrastructure inside their own borders is about control as much as performance. Data stays in the country, the model stays under local rules, and the bill stays with a domestic provider.
That is why the stack, rather than the model alone, is the real product. A model can be downloaded from anywhere. A full stack, built and operated at home, is a statement about independence.
HUMAIN has moved fast to build the pieces. Aramco completed a 49 percent acquisition of the company, and HUMAIN has signed SAR 2.5 billion in AI financing deals. A deployment using AMD Instinct MI355X accelerators and Cisco Silicon One networking went live in late August.
A Complicated Precedent
There is one more reason to watch December 15 closely. HUMAIN’s previous model launch came with a provenance problem.
In September, the company unveiled HUMAIN M3 at the LEAP conference in Riyadh, describing it as a frontier Arabic-language model. It later emerged that the model was built on MiniMax M3, the open-weight model from the Chinese lab MiniMax, and further trained on more than one trillion tokens of Arabic content.
HUMAIN said the Arabic training was real work, reporting an average score of 89.37 percent across seven Arabic-language benchmarks. But the episode showed that the line between building a model and adapting someone else’s can get blurry, and that the details matter.
That history does not prove anything about the new 3 trillion parameter model. It does explain why the AI community will want to see the weights, the architecture and the training details before celebrating.
HUMAIN+ and the Training Bet
The second announcement from the Riyadh event is easier to judge. HUMAIN+ is an applied AI education platform built with VUZ, a Riyadh-founded company that specialises in patented 360-degree video technology.
The centre was inaugurated with a ribbon-cutting at HUMAIN’s offices on November 11. It is due to become operational in the first quarter of 2027, offering about 1,500 learner places a year, with GPU-backed Builder Labs where participants can build prototypes on HUMAIN models and infrastructure.
Four learning journeys cover leadership, non-technical professionals, technical talent and students. Topics run from AI strategy and governance to product development, agentic AI, security and infrastructure operations. Assessment is competency-based, not just attendance and certificates.
Sector labs cover government, finance, healthcare, energy, education, tourism, logistics, legal and compliance, human resources and sport. An ALLAM-powered tutor supports self-paced study, and the platform combines HUMAIN Voice and HUMAIN Create.
Training a workforce is a slower, quieter bet than launching a giant model. It may also be the more durable one. A country that wants to run its own AI stack needs people who know how to build on it.
What Happens on December 15
December 15 is now the date that decides whether this story is a breakthrough or a bold claim. Three questions will answer it.
First, does the model actually appear, with its weights published? An open-weight model that stays behind an API is not the same as one the world can download and inspect.
Second, does HUMAIN publish benchmarks? Without independent numbers, nobody can judge whether 3 trillion parameters translate into better answers, or just a bigger download.
Third, who can use it, and on what terms? It is still unclear whether the model and the stack will be available outside Saudi Arabia, and at what price.
Amin’s own timeline is worth remembering. He told the Riyadh audience that the actual development time was about 15 months, from a company that has existed for roughly 18. That is fast by any standard, and speed is part of the story.
For now, the claim stands as a marker. Saudi Arabia has promised the world’s largest open-weight model, a full sovereign AI stack, and a date. On December 15, the world gets to check.
