Apple Chooses Qwen; TMTPost Analyzes Shift in Competition Among Domestic Large Models Toward the System Foundation

By | Shadow Memo

On October 31, the other shoe dropped. Apple officially announced its partnership with Alibaba, with Qwen formally integrated into Apple Intelligence, allowing Chinese users to directly invoke Qwen's text and image understanding and content generation capabilities in Siri, Writing Tools, and Visual Intelligence.

Developer Beta testing opened on November 4, and the feature will roll out to Chinese users with iOS 27.3 in December.

The news itself is not surprising. From Joseph Tsai confirming the partnership in Dubai in February 2025, to Apple Intelligence completing its filing in July this year, to Reuters reporting in August that the two sides were jointly training a China-exclusive model, this thread has been laid out for nearly two years.

What is truly worth dissecting is not the outcome—"it's finally settled"—but the hidden thread behind that outcome:

What Apple has been choosing was never the model with the highest parameters. It chose the foundation that would cause it the least trouble.

The impact of this on the competitive logic of domestic large models is far more profound than the launch of a system-level feature.

Eight in, one wins: the winner wins by "not looking like a rival"

According to cross-verification by multiple media outlets, over three years Apple approached eight vendors: Baidu, Alibaba, ByteDance, Tencent, DeepSeek, Baichuan, Moonshot AI, and Zhipu.

If ranked by benchmark scores and technical buzz, Alibaba may not have been the first choice. DeepSeek's impact in the open-source community and on inference costs is plain for all to see, ByteDance's Doubao leads by a wide margin in consumer user scale, and Baidu was the first major domestic tech company to proclaim "All in AI."

But the one that ultimately took the foundation seat was Qwen.

The real criteria of the elimination round are hidden in that list of "reasons for elimination." Baidu was initially held in the highest hopes, but its development progress did not meet Apple's standards, and the two sides differed on the boundaries of user data use.

DeepSeek and Zhipu, as startups, had sufficiently impressive model capabilities, but found it hard to match Apple's resource requirements for sustained delivery and support for hundreds of millions of concurrent users.

Where did Alibaba win? In one sentence: it is the one among the eight that is least "like a competitor." It doesn't make phones, doesn't make short videos, doesn't live off search advertising, and has no structural conflict with Alibaba's core e-commerce business or Apple's hardware ecosystem.

What Apple wants is a quiet background capability provider, not an ambitious player using Siri to funnel traffic into its own apps. Alibaba's business map is naturally suited to this role.

But that is only the first layer.

If this partnership is understood merely as "Apple chose Qwen's model capabilities," then the most crucial part is missed.

The problem Apple has faced in the Chinese market has never been finding a sufficiently smart large model, but finding a partner that can simultaneously solve three things: compliance, privacy, and engineering delivery.

Alibaba Cloud's domestic cloud qualifications, independently isolated computing clusters, and architecture in which data does not leave the country were Qwen's entry ticket.

A more critical step is hidden in the details: in the network access licenses for models such as the iPhone 18 Pro updated by the Ministry of Industry and Information Technology in October, the approved AI systems explicitly listed "Apple Intelligence, Alibaba Tongyi Qianwen, Tongyi Wanxiang."

The moment the model was written into the network access license, there was no longer any suspense at the regulatory level about the first wave of partners.

Privacy is Apple's red line. Apple's hard requirements are: user interaction data cannot be taken by partners to train their own models, cannot be retained long-term, and cannot be used for user profiling.

Alibaba's commitment is that Qwen user requests do not flow into the public model training pool, that a physically isolated independent tenant cluster is opened for Apple, and that the data link is closed-loop.

The line in the announcement—"user requests and generated responses will not be used by Alibaba for model training"—is not public-relations rhetoric; it is more like the baseline clause of the entire deal.

Alibaba Cloud's long-standing elastic scaling and fault redundancy capabilities for handling traffic peaks are a hard card beyond model benchmark scores.

Stack these three layers together and the conclusion is obvious: Qwen's victory is essentially the victory of "compliance infrastructure capability" over "pure model capability."

Apple cannot afford to wait; Alibaba's timing is just right

The precise alignment of timing is the most easily underestimated variable in this deal.

Apple's situation in the Chinese market no longer allows it to keep treating AI features as an option to be dealt with "later." In 2025, Huawei returned to first place in China's smartphone market with a 16.4% share, pushing Apple into second at 16.2%.

In the first quarter of 2026, iPhone sales in China fell 19% year on year, with market share dropping from 19.7% to 15.7% and its ranking falling to third.

When Huawei's Mate series can already tell its story with call summaries and AI photo editing, and Xiaomi and vivo have stuffed large models into photo albums and system tools, while "Apple Intelligence" in the settings page of China-market iPhones is still grayed out, this gap hurts high-end phone sales more directly than any spec comparison.

In July this year Apple Intelligence completed its filing, on September 29 it obtained regulatory approval, in October the network access license was settled, on October 31 came the official announcement, and in December comes the official rollout.

This is not a leisurely strategic choice, but a precise positioning driven jointly by competitive pressure and the regulatory window.

On Alibaba's side, the motive is equally clear.

Qwen does not lack technical buzz; what it lacks is a scenario where "users don't need to actively open the app."

In June 2026, the Qwen app had about 167 million monthly active users, while Doubao had 382 million in the same period and DeepSeek had 129 million.

Every Siri invocation, every writing assist, and every camera recognition by hundreds of millions of iPhone users is zero-cost reach.

What Alibaba is fighting for is not the ranking of an app, but the distribution path for the next generation of AI services.

The official narrative easily leads people to think that Qwen has taken over everything in Apple Intelligence for China. But code analysis and supply chain information point to a more finely structured arrangement: Alibaba plays the role of the "brain," providing core model capabilities for text and image understanding and content generation, covering iOS, iPadOS, macOS, and visionOS.

Baidu, meanwhile, is deeply embedded in Siri voice interaction and visual search modules, and the code identifier "Baidu Visual Search" has already appeared in the firmware.

Some reports say Qwen leads about 65% of tasks, while Baidu's Ernie handles the remaining roughly 35% of interaction and retrieval.

The significance of this division of labor is not who has more or less, but that it reveals Apple's real strategy: do not put all eggs in one basket, and do not let any single supplier have the ability to "choke the entrance."

Alibaba got the model foundation, Baidu got the interaction entrance, and the two are deliberately separated in the technology stack.

A deeper signal worth noting comes from a Reuters report in August: Apple is not simply calling Qwen's ready-made API, but with Alibaba's support has specially trained a customized model for the Chinese market, and the depth of cooperation goes far beyond "integrating a third-party model."

This means Alibaba's involvement is not only in inference calls, but also in supporting the training process.

From "fighting for the entrance" to "fighting for the foundation"

The strongest signal that the Qwen-Apple cooperation sends to the domestic large-model card table is not that "Alibaba won," but that the main battlefield of competition is shifting from user contention at the application layer to default capability positioning at the system layer.

On iResearch's annualized Token revenue ranking for August 2026, ByteDance ranked first with about 30 billion yuan, Alibaba second with 16 billion yuan, Zhipu third with 11 billion yuan, DeepSeek and Kimi at the 5 billion yuan level, and Baidu and Tencent at the bottom at the 3 billion yuan level.

ByteDance's lead in Token revenue is overwhelming, and Doubao's monthly active user scale is also far higher than Qwen's.

But Apple's choice of partner shows a cruel fact: Token revenue scale and consumer user volume do not automatically translate into eligibility for a system-level foundation.

What Apple looks at is compliance depth, identity neutrality, and certainty of engineering delivery—and these are precisely the things that cannot be made up in the short term through benchmark scores and DAU.

The implications differ for different players. ByteDance's Doubao remains the strongest presence at the application layer, but the "super app + AI assistant" model directly competes with Apple's system-level AI for users' time, which determines that it is very difficult for ByteDance to become the preferred foundation-type partner for hardware manufacturers like Apple.

Baidu has obtained the interaction entrances of Siri and visual search, which appears to have regained ground, but losing the primary position means its bargaining power at the model capability level has been compressed, and its role has retreated from "AI brain" to "functional module."

What DeepSeek and Kimi face is a more structural problem: model capabilities can be caught up through open source and algorithm iteration, but the time cost of compliance filing, the capital expenditure of cloud infrastructure, and the organizational ability to deliver to major clients cannot be solved by one round of financing.

Zhipu has already squeezed into the 10-billion-yuan club in Token revenue, but the risk discount that Apple-level clients apply to "startups" will be hard to eliminate in the short term.

More broadly, this marks a watershed: the first stage of large-model competition was "whose model is better to use," and the second stage is "who gets written into the operating system's default path."

The former compares benchmark scores and costs; the latter compares compliance, trust, and system-level delivery. The value of the ticket Qwen has obtained does not lie in how much API call revenue it can bring in the short term, but in the fact that it has nailed the perception that "Alibaba = the foundation of Apple AI in China" into the procurement lists of all hardware manufacturers.

Going forward, when any phone maker, PC maker, or car-machine maker chooses an AI partner for China, Apple's choice will become an unavoidable reference point.

Apple chose Alibaba mainly because it believes that, before the three thresholds of compliance, privacy, and engineering delivery, Qwen is the choice least likely to make Apple Intelligence late in China once again.

Perhaps in the future business module of large models: the standard of competition has, from this moment on, been permanently raised by a notch.