Regional AI Models: One Product, Several Systems

Apple Intelligence cleared its final regulatory hurdle in China this week. The version Chinese users will get is not the version on your phone.

Same name. Same icon. A different model underneath.

That gap deserves a name, because regional AI models are about to become a documentation problem for anyone running a global AI product across borders.

One approval, one revealing detail

China’s cyberspace regulator said on Wednesday that Apple Intelligence has been registered for use on iPhones in China. Registration is no formality there. China requires companies to register large language models and generative AI services with regulators before they reach the public, so without it the service does not ship.

The revealing detail sits in what was registered. Alibaba told Reuters that its Qwen model will be integrated into Apple Intelligence across iOS, iPadOS, macOS and visionOS for users in China. A Baidu spokesperson confirmed that Apple is working with Baidu on features for Chinese iPhone users. The regulator gave no launch date.

One brand, then, across every market. A different engine in one of them.

Regional AI models are a design pattern now

Regional AI models are what happens when a global product meets a national rulebook. A market demands registration, or data localisation, or a domestic partner. The vendor complies by swapping the component that is easiest to swap, and that component is the model. The interface survives. So does the brand. What changes is the system.

Apple is simply the largest example to surface this month. The pattern belongs to any vendor that wants access to a market running its own approval regime.

What regional AI models do to your documentation

Here is where it stops being trivia. Governance obligations attach to systems, not to logos. The EU AI Act asks providers and deployers to know what they have put into service: what it does, how it was built, where its limits sit, what risks it carries. Every one of those answers moves when the model moves.

So if your vendor ships regional AI models, your documentation inherits the split. A single entry describing “the vendor’s assistant” stops being accurate. It now describes a system your users in one region have and your users in another do not.

Consider what actually differs across regional AI models. Training data differs. Content filtering differs, sometimes for reasons political rather than technical. Behaviour under an identical prompt differs. Where inference happens differs, which drags data residency into the picture. The approval that permitted the deployment differs, and so does the regulator standing behind it.

None of that is visible from the interface. All of it is in scope for the questions a supervisory authority might put to you.

Keeping a register of regional AI models

The practical response is unglamorous and cheap: know what you have, per region. A register of regional AI models can be a short table kept beside the vendor contract, recording for each market:

• the product name as users see it
• the underlying model and its provider
• the regulatory approval or registration that permitted it
• where inference and data processing occur
• known behavioural or content differences from other regions
• the date the vendor last notified you of a model change

That final line does most of the work. A vendor who cannot say when the model behind the brand last changed is a vendor whose documentation will not hold up.

Notification is the part worth negotiating early. A model swap can arrive as a routine update, announced to nobody, and it can shift behaviour your risk assessment was written against. Contract language that obliges the vendor to tell you when regional AI models change, and how, converts an invisible event into a governed one.

The layer underneath: contradictory obligations

Now the awkward part. On 8 June, the Pentagon added Alibaba and Baidu to its Section 1260H list of Chinese military companies. The designation is not a sanction. It bars the US Defense Department from contracting directly with listed firms from the end of June, and from procuring their products or services through third parties from June 2027. Alibaba rejects the listing and has said it will take all available legal action.

Set the two facts side by side. Beijing approved an American company’s AI product on the practical condition that it runs on Chinese models. Washington designated the providers of those models five weeks earlier.

Neither government did anything irrational. Each applied its own rulebook. Yet the result is that one vendor relationship now carries obligations pointing in opposite directions, depending on which jurisdiction is asking.

For an EU organisation, that is no abstract tension. It is a question about whether your AI supply chain can satisfy two regimes at once, and what happens when it cannot.

This is not an Apple story

The specifics will age. The pattern will not.

Any vendor with global ambitions and a Chinese market faces the choice Apple faced. The same holds wherever a state ties market access to local approval and local models, and India and the Gulf states are building approval regimes of their own. Regional AI models are the predictable consequence of that arithmetic, not an exception to it.

For a governance function, the shift is quiet but real. The unit you govern is not the product. It is the product in a jurisdiction, running a specific model, under a specific approval. That means more rows in a register and more questions for a vendor, which is considerably cheaper than discovering the split during an audit.

Firms that already track where the rulebooks diverge will find regional AI models a small extension of work they do anyway. Future Prep keeps mapping those divergences as they land.