We’ve made it, everyone!
China’s large models have gone global!
Over the past couple of days, a major scandal shocked everyone.
Recently, a Saudi netizen proudly announced: We have launched HUMAIN M3, a national-level large model with 100% Saudi heritage.
It surpasses the world's most powerful AI models, perfectly aligns with Arab culture and language, is deployed on local servers in Saudi Arabia, and will soon open its weights!
Key point: This company has an extremely prestigious background, backed by Saudi Arabia’s Public Investment Fund (PIF), with the Crown Prince of Saudi Arabia serving as its chairman.
Wow, Saudi netizens are thrilled— the comments section is buzzing.
What was the result? Before the post even made it through the night, several researchers on X spotted the issue.
A mixture-of-experts model with 428 billion total parameters and 23 billion activated parameters... hmm? This recipe seems increasingly familiar.
Soon, a brilliant comment appeared: “100% Saudi? Aren’t you afraid of Allah’s punishment?”
The publisher, seeing that the deception could no longer be hidden, had to personally clarify: In fact, the underlying core of HUMAIN M3 is based on a Chinese AI company MiniMax the M3 model.
Today I saw the newly released official press release — apparently HUMAIN has long been in China MiniMax strategic partner!
This is not a private shell company—it’s the national AI showcase under Saudi Arabia’s sovereign wealth fund, which has issued real checks to become a major corporate client of China’s large models!
Wow, the Middle Eastern billionaires invested heavily to build a national showcase, yet inside it beats a "Chinese heart"!
China's large models are now beginning to serve as others' national infrastructure.
Don't underestimate the ambitions of the Saudis.
You might think this is just a wrapper.
Thinking this way underestimates the ambition of the Saudis.
Keep in mind that Arabic is one of the most challenging languages in the world.
It slants from right to left, has a complex structure, and exhibits severe dialectal variation—such as Egyptian Arabic, Gulf Arabic, and Levantine Arabic—making them almost like different languages.
The Arabic in "Welcome to Dragon Restaurant" sounds more similar to the widely spoken Egyptian Arabic dialect.
Moreover, high-quality Arabic language corpora have long been scarce on the internet.
Using a general-purpose U.S. large model to process Arabic will always carry a strong machine-translation flavor.
How do Saudis do it?
They took it directly. MiniMax M3, this powerful Chinese foundation, was fine-tuned by feeding it over one trillion pure Arabic tokens.
This trillion-word corpus fills not only linguistic gaps but also captures authentic colloquialisms, local culture, religious practices, and even thought patterns of the Arab region.
As a result, sheer force brings about miracles.
HUMAIN M3 achieved the highest scores in five out of seven Arabic benchmarks, with an equal-weighted average score of 89.37%.
This score utterly dominates top proprietary models like GPT-5.6 Sol and Claude Opus 5.
In the Arabic language vertical, it outperforms the original base model by an average of 9 percentage points, achieving true leadership.
This is no longer a model that speaks Arabic; it is a model built specifically for Arabic.
Moreover, uniquely, it is natively multimodal from birth—able to understand images and text, long videos, and even interact directly with computer screens, making it purpose-built for future AI agent workflows.
Chinese models are becoming the overseas AI infrastructure.
Actually, the significance of this partnership lies in that it represents a new form of Chinese AI going global.
In the past, Chinese AI companies going global either bought traffic for their apps overseas or sold their APIs to overseas developers on a per-call basis.
Although you can make quick money, your fate is always in someone else’s hands and you could be replaced at any time.
But this time, MiniMax The cooperation with Saudi Arabia represents an entirely new approach: China's model is becoming the technological foundation for other countries' AI systems!
This brings to mind Huawei back in the day.
Previously, Chinese smartphone manufacturers going overseas simply sold a few low-cost phones. But Huawei’s overseas expansion involved helping telecom operators in the Middle East, Africa, and Europe build 4G and 5G communication base stations and lay submarine fiber-optic cables.
Selling phones is business; building base stations is about creating an ecosystem and setting the rules.
Current MiniMax We're building the overseas infrastructure for the AI era.
Perhaps in the future, many other countries' domestic large models may be built on MiniMax Qwen, DeepSeek GLM. This is the true outward expansion of China's open-source large models.
Excitingly, the HUMAIN platform will become Saudi Arabia’s national AI infrastructure, connecting Saudi developers, researchers, and businesses.
This means that most of Saudi Arabia’s future AI applications—autonomous driving, government systems, and medical AI—are likely to be built on this Chinese foundation.
This marks the first time a Chinese large model has truly entered the core circle of global sovereign AI.
The world has long suffered from closed-source systems.
Why is Saudi Arabia going to such great lengths to develop its own sovereign large model, even bringing in the Crown Prince to lead the effort?
In the AI era, data is oil, and models are refineries.
As a major Middle Eastern country heavily reliant on energy transition, Saudi Arabia cannot transmit its core data—government, financial, medical, and cultural—to servers owned by OpenAI or Anthropic in California.
If one day a ban shuts down accounts, the entire country’s AI ecosystem would immediately collapse.
Therefore, "sovereign AI" has become a necessity for non-English-speaking countries worldwide.
To this end, the large model must meet the conditions of being deployable on servers within the country, embodying local cultural values, and being fully autonomous and controllable.
At this point, China's open-source large models become the best allies in the world.
Hugging Face's latest report shows that models developed in China accounted for 41% of platform downloads over the past year, historically surpassing the United States for the first time!
Big tech companies in Silicon Valley aim to create "just one strongest model in the world."
Meanwhile, Chinese large models are pursuing another path: in addition to becoming the world’s strongest models, we provide the best open-source foundation—you take it, and build your own models by combining them with your own language, culture, and data.
Remember, a weak foundation leads to instability. If the base model lacks state-of-the-art capabilities, no subsequent training can achieve a breakthrough.
It has been proven that the paradigm of "strong foundation + localized deep fine-tuning" far outperforms reliance on English-dominated general-purpose closed-source models for non-English-speaking regions.
China's open-source foundation plus localized fine-tuning is the perfect approach.
HUMAIN M3's comprehensive superiority over GPT and Claude in various evaluations has already demonstrated this.
Mutual effort, crossing mountains and seas
Interestingly, if you've been following recent major national initiatives, you'll realize this partnership is no coincidence.
Just recently at WAIC, senior officials explicitly stated the need to "support global AI capacity building" and even specifically called for the establishment of an international AI application cooperation center aimed at the Arab League.
The emergence of HUMAIN M3 aligns perfectly with this manifesto.
Driven by the Vision 2030 initiative, Saudi Arabia is making a national bet on AI, with reports indicating plans to invest up to $100 billion to build a national AI giant.
Saudi Arabia doesn’t lack money, but its anxiety over "digital colonization" cannot be resolved by U.S. giants—American AI models like Fable are at constant risk of censorship and supply cutoffs.
More critically, core data from the Middle East, such as government and energy records, must be uploaded to the cloud, directly violating local legal requirements.
What China's large models provide is the "digital security" that Saudi Arabia most urgently needs.
The West only wants to sell you licensing software, but China is willing to help them build their "blood-producing system" in their own data centers in Saudi Arabia.
Saudi Arabia has capital, computing power, and vast amounts of Arabic-language data; China has cutting-edge foundational model architectures like 428B MoE, mature training engineering expertise, and a culture of open and inclusive ecosystem sharing.
Two ancient civilizations intersecting in the age of AI.
China's large models are becoming like "water, electricity, and gas," powering the servers of Middle Eastern billionaires and nurturing a thriving Arabic-language AI agent ecosystem.
China's large models become the "shipbuilders"
Looking back over the past few decades, we have always seemed to be followers.
In the PC era, the foundation was Windows, and we developed software; in the mobile internet era, the foundation is iOS and Android, and we develop apps.
For a long time, even China’s most profitable internet companies operated under rules set by others.
But today, the script has finally been rewritten.
China's large models are beginning to serve as other countries' "national infrastructure."
From MiniMax From M3 to HUMAIN, a technological legacy spanning thousands of miles. In the global open-source AI ecosystem, we have reached the top of the food chain.
Today, hundreds of millions of people speak Arabic worldwide, and hundreds of millions more will speak Hindi, Spanish, and Swahili in the future.
They will find that the most open and user-friendly foundation bears these words: Great AI, Made in China.
The vast ocean of Chinese open-source models has just begun.
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.