Apple has trained its own large language model for the Chinese market, Reuters reported Friday, citing three people familiar with the unannounced work. Alibaba supported the training. The company had been expected to do something considerably simpler: license a domestic Chinese model and wire it into Apple Intelligence the way it wired in ChatGPT everywhere else.
Instead, according to the report, Apple is running a dual-track setup — its own China-specific model alongside Alibaba’s Qwen and technology from Baidu. That makes Apple the first foreign company cleared by Chinese regulators to offer a proprietary AI model inside the country. Apple Intelligence is expected to arrive on Chinese iPhones in the coming months via an iOS update. Apple has not commented publicly, and the architecture — specifically, which model actually answers when a Chinese user asks Siri something — has not been disclosed.
The signal was there five days earlier. On August 10 Apple briefly published a support guide in China explaining how Mac users could connect Qwen to Siri and Writing Tools, then pulled it in under a day. At the time that read like a slip. It now reads like one track of a two-track plan surfacing before the other was ready.
Why build rather than borrow
Licensing was always the cheap path, and Apple has spent eighteen months looking like it would take it. Building instead buys three things. First, control of the content layer: a model Apple trained is a model Apple can tune to satisfy Chinese content rules without handing the entire user experience to a partner. Second, a hedge — if the regulatory relationship with any single Chinese lab sours, the feature does not go dark. Third, and least discussed, it keeps Apple’s privacy story roughly intact in a market where the alternative is routing every Siri query into someone else’s infrastructure.
The cost is that Apple now maintains a separate frontier-adjacent model for one country, with the compute bill and the update cadence that implies. That is not a small ongoing commitment for a company that has repeatedly signalled it would rather partner than train.
Our take: The headline is regulatory, not technical. “First foreign company approved to run its own model in China” is a permission slip nobody else has, and it arrives while Apple is losing share to Huawei on the strength of AI features Chinese buyers can actually use today. The real read-through for everyone else: in the largest markets, AI is stopping being a feature you license and starting to be a jurisdiction you have to be licensed for. If your product has an AI layer and a China, EU or India footprint, the question is no longer “which model do we call?” but “how many models do we have to maintain, and who signs off on each one?” That is a cost line, not an engineering detail. And keep the caveat attached — this is reported, unannounced work from anonymous sources. Apple has said nothing.
What to watch
- Which model answers Siri. The routing logic between Apple’s own model, Qwen and Baidu is the whole product, and none of it has been described.
- The iOS update that ships it. A date turns a Reuters report into a shipped feature; until then this is sourcing, not launch.
- Whether the approval is genuinely a precedent or a one-off. If a second foreign company gets the same clearance, the market just changed shape.
- Huawei’s response. Apple is arriving late into a market where domestic handsets have shipped AI assistants for over a year.
- Whether Apple extends the build-your-own approach beyond China, or treats it as a compliance exception.