Apple Trained Its Own AI Model for China, 22 Months After Promising Apple Intelligence There
Three people told Reuters that Alibaba helped train it. The bigger claim in the story is that Beijing has never cleared a foreign company to run its own AI model inside China.
Apple has trained a large language model built only for the China market. Three people familiar with the work told Reuters, which published the story on 14 August. Alibaba helped train the model, the same people said. They asked not to be named because the work is sensitive and not public.
This is a change of plan for Apple in China. The approach there had been to borrow. Chinese iPhones were set to run Apple Intelligence on models from Alibaba and Baidu, because Anthropic's Claude and OpenAI's ChatGPT are not available in the country. Apple and Alibaba did not reply to Reuters when asked to comment.
Reuters also says the model would make Apple the first foreign company Beijing has approved to offer its own AI model in China. That line is the one worth sitting with. Training a model is ordinary work for Apple. Getting permission to run one in China is not.
The 22-Month Wait That Ended on 15 July
Apple began promising Apple Intelligence in China when the iPhone 16 launched in September 2024. Every mention carried the same footnote: subject to regulatory approval. The filing took nearly two years to clear.
On 15 July 2026, the Cyberspace Administration of China published a registration list covering seven on-device generative AI services for phones. Apple Intelligence was on it, filed under Apple Technology Development (Shanghai) Co., Ltd. The other six entries went to handset makers already selling in China, Huawei among them.
The registration set no launch date. What it did fix was the split between two Chinese partners. Alibaba's Qwen handles language, and Baidu handles visual and AI search. Apple had picked Baidu as its lead China partner first. Alibaba took over that role by February 2025, after Baidu's models fell short of what Apple wanted.
An Alibaba spokesperson told CNBC what the deal covers.
"Qwen will be integrated into Apple Intelligence experiences within iOS, iPadOS, macOS, and visionOS for users in China."
Alibaba spokesperson, to CNBC
Alibaba's US-listed shares rose about 4% in premarket trading that day. Reuters now says Apple Intelligence should reach China in the coming months, after an iOS update.
Apple Rents a Model at Home and Owns One in China
Look at what Apple did in its own market over the same stretch. In January 2026, Apple and Google confirmed in a joint statement that Gemini would power the rebuilt Siri. Apple said Gemini "provides the most capable foundation" for its AI. Bloomberg had reported the price at roughly $1 billion a year, for a model of 1.2 trillion parameters.
So in the market where Apple can pick any supplier it wants, it paid a rival. In the market with the tightest rules and the fewest options, it trained its own. The usual telling of Apple's AI story is a company that fell behind and had to rent. China turns that around.
There is a plain business reason for it. A model Apple trains is a model Apple can hold steady. Partner models get updated, repriced or pulled, and Apple has no say in any of that. Reuters describes the result as a dual-track setup, with Apple's own model working next to approved Chinese ones. No other foreign company in China has that arrangement.
It also gives Apple somewhere to stand if a partnership sours. Apple has spent this year in court with OpenAI over trade secrets, and it went into that fight with a Google deal already signed. Owning the China model means one fewer relationship that can take the product down with it.
Apple's Qwen Guide for Macs Lasted One Day
Six days before the Reuters story, Apple published a Chinese-language support page for Macs. It explained how someone in mainland China could connect Qwen to Siri and to Writing Tools. Siri could hand Qwen requests that needed deeper reading of photos or documents. One example had a user ask Siri to get Qwen to write a poem.
The extension needed macOS 26.6 or later and the user's own Qwen account. Alibaba was barred from training on anything users sent through it. Apple published nothing like it for the iPhone or iPad, even though Alibaba has said Qwen will cover all four of Apple's operating systems in China.
The page went up on 8 August and was gone by 9 August. Apple gave no reason.
Apple Support told AppleInsider that it had heard nothing about a launch. This is the translated statement.
"We are notified in advance whenever a feature or new project is launched. We have not received any such notification, and the feature integrating Apple Intelligence with Qianwen [aka Qwen] has not yet been released in mainland China."
Apple Support, via AppleInsider
Liu Dingding, a veteran tech analyst, gave the Global Times another reading on 9 August. The takedown "might signal pending tweaks to technical details amid ongoing regulatory progress", he said.
Read that page again knowing Apple has a China model of its own. A guide describing only a Qwen extension would be an incomplete picture of what Apple is shipping. Apple has explained nothing, so this is a guess and should be treated as one. The timing does fit. It was also the second Apple AI feature to appear early in China this year. Apple Intelligence reached some Chinese users by mistake in March 2026 before it had approval, and Apple pulled it.
Huawei at 22.6% and Apple at 18.1% in the Same Quarter
Reuters frames the model as an answer to losing ground to Huawei. Across the last few years that is fair. In the most recent measured quarter it is not what the numbers show.
IDC put China's smartphone market down 4.3% in the second quarter of 2026, at roughly 66 million units. That was the fifth quarterly fall in a row. Huawei led on 22.6% share, up from 18.1% a year earlier. Apple came second on 18.1%, up from 13.9%.
Apple's shipments grew 24.4% year over year, the strongest growth of any major brand in the market. Its Greater China revenue came in at $18.82 billion for the June quarter, up 22% from $15.37 billion. IDC credited stable pricing and heavy promotions, plus buyers moving early ahead of expected iPhone 17 price rises.
Apple grew that fast in China with no Apple Intelligence at all. That weakens the idea that missing AI features were the main drag on sales. It also raises the bar for the new model. Apple is shipping AI into a market where it is already growing, which is a harder place to prove a feature earned its keep.
Beijing Was Weighing Curbs on Qwen Five Weeks Ago
Apple's partner has its own political weather. On 7 July, Reuters reported that Chinese commerce officials had met ByteDance, Alibaba and Z AI to discuss limits on foreign use of their best models. Qwen was named in those talks. The discussions covered closed and open models, funding limits for startups, and penalties for leaking proprietary AI.
That was five weeks before this story, and it came a month after a second problem for Alibaba. Anthropic told the Senate Banking Committee in June that Alibaba had run the largest distillation attack it had seen. The method means querying a frontier model at scale, then using the answers to improve a weaker one. Anthropic asked Washington for sanctions against the Chinese labs behind it.
Neither item touches Apple directly. Both describe a partner whose model access can move for reasons Apple has no control over. Washington ran the same play in June, with export controls on Anthropic's Fable and Mythos models. If your China AI experience rides on one company's model, training your own starts to look like insurance.
"With Alibaba's Support" Covers Three Different Things
Reuters is careful about the edges of what it knows. Its closing line puts the main gap plainly.
"It was not immediately clear what role Apple's self-trained model would play alongside third-party Chinese models and how those are working together."
Reuters, 14 August 2026
The phrase doing the most work in the story is "with Alibaba's support". It could mean Alibaba rented Apple the compute to train on. It could mean Alibaba supplied Chinese-language training data, which is the part Apple is weakest on. It could also mean Apple started from Qwen weights and kept going. Each reading leaves Apple with a different amount of independence, and the last one would make Apple's model a Qwen derivative. That is a much smaller step than the headline suggests.
There is a second gap in the framing. Apple already trains foundation models. It has shipped an on-device model and a server model under Apple Intelligence since 2024, and Bloomberg put the server model at about 150 billion parameters. Apple's June 2026 Siri runs on Apple's own models built with Google, not on the Gemini that Google ships to its own users. A China-specific model is not new ground for Apple's engineers.
The new ground is regulatory. Under China's rules, a public generative AI service has to be registered, and foreign firms have in practice needed a local partner. Apple's July filing sits under a Shanghai subsidiary and describes an on-device service. A proprietary Apple model, trained in part with a Chinese company, is a case the CAC has not published a precedent for.
So watch the register when the iOS update lands. Benchmarks and demos will arrive first, and they will tell you less than one line in a government list. What matters is whether a second Apple filing appears, and whose name is on it.
