Apple Chooses Qwen for Siri Integration, Alibaba AI Achieves Major Result

Apple Chooses Qwen, Alibaba AI Scores a Big Win

Wang Zhiyuan Download the app for exclusive early access October 31, 2026, 13:35:18, from Beijing

When I got up this morning, I opened Apple's website. I had originally wanted to check out the newly released iPad mini.

Apple says Qwen has been integrated into Siri, Writing Tools, and Visual Intelligence, and will officially roll out with iOS 27.3 in December.

...

I went back and looked at the filing information from that time.

The filing was approved on July 8, and the Cyberspace Administration officially announced it on July 15. The registered name is "Apple Intelligence," and the filing entity is Apple Technology Development (Shanghai) Co., Ltd.

There were seven products in the same batch. Huawei's Xiaoyi, OPPO, vivo, Xiaomi, Samsung, and Nubia were all on the list. Apple was wedged in among them and did not look conspicuous at all.

But Apple had waited more than two years for this certificate.

As early as 2023, there were reports that Apple had begun looking for AI partners in China. It searched very cautiously, and lists circulated round after round, but Apple officially never acknowledged any of them.

That was true until February 13, 2025, at the World Governments Summit in Dubai, when Joseph Tsai was asked about it and did not beat around the bush. He said:

Apple needs a localized partner in China to serve its phones.

Apple has always been very picky. They talked with many companies in China; in the end, they chose to do business with us. We are very lucky and very honored to do business with a great company like Apple.

Note that the words "very picky" were Tsai's own. On the day that quote came out, Alibaba's Hong Kong shares rose 9%, hitting a three-year high.

From that day to this announcement on October 31, 20 months passed.

I think there is one thing worth pondering. It got the filing in July, and logically it could have moved forward then, but Apple did not act until it officially issued the announcement at the end of October.

During those three months, it was most likely still doing adaptation, testing, and some communication with regulators.

Have you ever thought about it? After picking for a full two years, what exactly was Apple picking? I think it may not have that much to do with model benchmark scores. What it really had to pass was probably three tests.

First, the capabilities had to be complete enough.

Q&A in Siri, rewriting and generation in Writing Tools, and image description in Visual Intelligence are three completely different jobs. Text understanding, image understanding, content generation, and image generation—none of them can be missing. Being strong in just one modality probably would not be enough.

On Qwen's side, the flagship model is Qwen3.8-Max, with 2.4 trillion parameters, native support for a 1-million-token context, and only 95B activated per token.

Those figures are absolutely in the global top tier, but for Apple, another thing may have been more useful.

Qwen has everything from small on-device models to cloud-based flagships, in all sizes, so there is no need to cobble things together. It has already open-sourced more than 460 models, with global downloads exceeding 3 billion.

There is also a third-party report.

In September this year, Gartner released its first Magic Quadrant for Generative AI Model Providers, and Alibaba Cloud was the only Chinese company to enter the Leaders quadrant.

But all of that together may only have been enough to get Apple to put you on the candidate list.

What was truly difficult was the second test: whether it could be embedded.

Apple's setup does not require a separate app. It is in Siri, in Writing Tools, and in the camera viewfinder. When a user taps, they have no idea who is being called behind the scenes, and they do not particularly care.

There is a detail here that I think explains the issue better than parameters do.

Apple named a feature in the Chinese version of Writing Tools "Qwen Creation"; in English, it is Compose with Qwen. The overseas version is called Compose with ChatGPT.

I compared them. The sentence structure is exactly the same; only the name was changed.

This shows that Apple reserved the same slot, but in China it had to switch to a local model and also let users know who is doing the work.

There is also the account.

Users can decide for themselves whether to enable the Qwen integration. When enabling it, they can either sign in with Apple or directly use a Qwen app account. After signing in, it can be called from Siri, Writing Tools, and Visual Intelligence.

When I looked through both sides' press releases, I also saw a promise: Alibaba will not use users' requests and generated replies to train models.

On the surface, that sentence is written for users, but it is actually also meant for regulators.

Third, it had to hold up.

Exactly how many devices in China can run this setup? No one has given an accurate figure so far, but it is certainly in the hundreds of millions.

Writing Tools is a system-level feature, and Siri can be awakened by users at any time. Concurrency at that scale is completely different from building a chat app.

Privacy is another hurdle.

Overseas, Apple talks about private cloud computing, where data does not enter its own servers. In China, the architecture still has to be realigned with local cloud services and local compliance requirements.

To meet regulatory requirements, Apple also had to work with local partners. In the July filing, the entity was under Apple's Shanghai company, but for the model, what passed review was Qwen.

After passing those three tests, the October 31 announcement came.

However, one thing needs to be made clear: what is being rolled out this time is not yet the full set. I read the explanation on Apple's website, and it is very clear:

This update does not include Siri AI, nor does it include the batch of next-generation features announced at this year's WWDC. Support for other platforms still has to wait for the relevant approvals to be completed.

Quite a few features have been confirmed so far.

Real-time translation, writing tools, visual intelligence, the erase tool in photos, Genmoji, natural language search in photos, memory movies.

Priority email, email summaries, smart replies, notification summaries, priority notifications, reduce interruptions, audio transcription summaries for notes, call transcription summaries, and talking to Siri by typing.

Looking at this long list, there really are quite a few features. But a careful comparison reveals that the vast majority of them were already launched in the US two years ago.

So what's being brought over this time is actually the version that was owed before; the most advanced set is still to come.

......

After all the sifting, few Chinese local companies could enter Apple's field of vision, but for such an important partner, after repeated deliberation, Apple still chose Alibaba. I think this is because of two unique advantages Alibaba has.

First, in China, Alibaba's comprehensive B-end customer service capability is top-tier.

I looked back and found that this kind of work Apple needs is really not the first time Alibaba has done it. The Olympics is a pretty good example.

In 2017, Alibaba became a global partner of the International Olympic Committee, and in 2018, it worked with Olympic Broadcasting Services to build OBS Live Cloud.

At the 2021 Tokyo Olympics, this cloud broadcasting system was put into use for the first time, and it was also the first time in Olympic history that cloud computing supported global broadcasting.

In the years since, it gradually went from an option to a core distribution platform; at the 2024 Paris Olympics, remote broadcasting relied mainly on it. By the 2026 Milan Winter Olympics, it will need to support 39 broadcasting organizations and 428 live signals, 26 of which are ultra-high-definition.

However, what is truly similar to Apple in the Olympics matter is another thing.

OBS is building an automated media description system, and what powers it is Qwen. It can identify athletes and key moments in competitions within seconds, automatically generate event descriptions, and conveniently tag videos.

There is also an Olympic image library based on Qwen, storing 8 PB of historical material; staff only need to enter a sentence like "retrieve the full performance of the figure skating gold medalist" and can find it almost immediately.

If you put these two things together with Apple's cooperation this time, you'll find they are very similar: both involve installing components into someone else's system, and once installed, they are hidden away, so users cannot feel them.

This kind of work is difficult because it requires grinding along with someone else's system. Whether the model is smart enough is of course important, but being smart alone is not enough.

On the Olympics side, Alibaba and OBS have been grinding from 2018 to now, eight years. On Apple's side, it also needs to find a partner that can be polished and iterated over the long term.

Second, Alibaba already has a complete set of things in hand.

From chips to models, and then to the cloud, it is basically all there. On the chip side, Pingtouge's AI chip Zhenwu has already been deployed at large scale on Alibaba Cloud. The Zhenwu V900 released in September can scale a single cluster to 500,000 cards.

On the model side, in addition to the flagship with 2.4 trillion parameters, there are reports that Alibaba is training a new model with 5 trillion to 10 trillion parameters.

On the cloud side, Alibaba Cloud currently has 106 availability zones in 31 regions globally, and plans to expand its data center scale to more than 20 GW by 2032.

For Apple, these things may be more substantial than mere model parameters.

When Apple talks about private cloud computing overseas, the core is to keep data processing within its own security boundary; Alibaba has a product called Apsara Stack, which can move public cloud capabilities into customers' own data centers, with both training and inference deployable on demand.

It can be privatized and also deployed in a hybrid way, so at least architecturally, it is easier to align with Apple's requirements.

How much is that certificate really worth?

I also looked at the data IDC just released at the end of October; in the first half of 2026, Alibaba Cloud took a 46.6% share of China's large language model MaaS market. Second place Zhipu had 10.6%, and third place Volcano Engine had 10.2%.

The entire market size was 10.63 billion yuan, up 889% year on year. This business is growing fast enough, right? But why should customers trust you?

Apple's order happens to answer this question.

A company famous for being picky ultimately handed over a system-level entrance and let Qwen into Siri.

I think the persuasiveness of this matter is stronger than holding ten launch events; in the future, when Alibaba Cloud goes to negotiate with major customers and they ask, "What projects have you done?", Apple is a ready-made answer.

So the truly valuable part of this deal probably has to be found in cloud and MaaS orders.

Looking back, Apple wanted to find a partner that could be embedded in its own system and was reliable and stable. What Qwen took away was exactly that "qualification certificate" officially stamped by Apple.

Yeah, both sides got a big result from this deal.

"Special statement: The above content (including videos, pictures, and audio if any) is uploaded and published by users of "Dafeng Hao", a self-media platform under Phoenix Net. This platform only provides information storage space services.

Notice: The content above (including the videos, pictures and audios if any) is uploaded and posted by the user of Dafeng Hao, which is a social media platform and merely provides information storage space services."