Hands-on test of Apple Intelligence in China: on-device features filled in, complex tasks still need Qwen as a fallback

Although I had long expected a “castrated version,” Smart Things couldn’t wait and stayed up all night to try it out.

Those who want to try it now can upgrade to iOS 27.3 Beta, enable “Apple Intelligence & Siri” in Settings, and experience it once Siri finishes updating in the background; the official version will roll out to all iPhone users in China in December alongside iOS 27.3.

First, the image: this table shows all the capabilities in this Apple Intelligence update:

▲Statistics on Apple Intelligence features launched in China (compiled by Smart Things based on Apple’s official website)

After the full experience, my biggest takeaway is that there is not much surprise; most of the capabilities this Apple Intelligence suite focuses on have long been maturely implemented by domestic AI phones. But Apple’s hybrid Apple + Qwen approach gives users greater autonomy, with users deciding throughout whether to enable Qwen capability enhancements.

The first batch of Apple Intelligence features can be divided into two categories. One mainly consists of efficiency features that help users save time and reduce information burden, such as using AI to filter emails, messages, and notifications for you; the other focuses on generative AI capabilities, such as rewriting and proofreading text, summarizing audio transcriptions, photo erase, and real-time translation.

Moreover, the linkage logic between Apple AI and Qwen is very smooth. On-device native features focus on capabilities such as translation, transcription, and photo processing. When encountering complex creation, in-depth Q&A, and other tasks it cannot handle itself, it will proactively invoke Qwen, and before the first call it will always ask for the user’s consent.

For example, when I ask Siri, “I’ve been sleeping poorly lately—are there any ways to improve it?”, it proactively asks whether to use Qwen to handle it. After getting the answer, I can tap the Qwen logo in the lower left corner to jump to the Qwen app for a more complete answer.

Apple Intelligence was officially announced at WWDC in 2024 and rolled out to overseas users in October of the same year. It has now been more than two years since it officially launched in China. By contrast, domestic peers Huawei, Xiaomi, OPPO, and vivo had already centrally upgraded their phone assistants into a new generation of AI assistants powered by large models as early as 2023 and 2024, and now they are already competing on Agent capabilities.

But Apple Intelligence currently cannot autonomously execute complex long tasks, nor can it temporarily invoke other vertical Agents.

By the way, a small aside: just before this Apple AI update, the Apple Weather app had a widely circulated blunder, forecasting a minimum temperature of -31°C next Friday. This bug has now been fixed.

Real-time call translation, photo erase, Genmoji, photo search, automatic memory movie generation, notification summaries, strong reminders in Focus mode, and call audio transcription are features natively available in Apple’s on-device models.

Real-time translation is one of the highlight features. Now we can translate in real time when sending messages and making calls, and during calls we can also record, transcribe, and summarize. For example, when I was on a call with a friend who was speaking German, I turned on real-time translation, and the AI amplified the Chinese translation in a human voice while lowering the original voice of my friend speaking.

Currently, real-time translation on iPhone supports 9 languages. One small reminder: before using each language, you need to download the corresponding language pack first, each about 800MB, which still takes quite a bit of time, so it is best to download them in advance.

In addition to real-time translation, we can now directly record and transcribe during calls. After the call recording ends, the audio file, transcribed text, and summary will appear in Notes. Because the speaker’s voice is relatively low during phone recording and there is noise interference, the generated text is intermittent, but the overall summary is quite good, including important elements such as “raining, taking the kids out to play.”

However, I tried the translation feature in the Messages app and it never translated successfully.

Genmoji focuses on DIY emoji packs. We can mix and match existing emoji to create new expressions, and we can also add photos of ourselves or friends to generate emoji or prototypes for new Genmoji. This will make it easier to make memes for friends in the future.

You can also directly describe your needs in text. For example, if you enter “I have to go out on a rainy day, so cold and miserable,” it will automatically combine the following emoji. The expressions it generates are simple and direct: a rainy background + a miserable expression, or ice cubes around the expression to indicate cold. But compared with mature tools like WeChat emoji, Genmoji’s available creative space is relatively thin.

The Photos app has also caught the AI train. We can search for photos by text description and generate memory movies with one tap. I entered keywords such as “seaside,” “last year’s yellow kitten,” and “last week’s gray kitten,” and it matched the corresponding materials.

Photo erase is also a very practical feature. iPhone’s erase feature requires me to manually circle the part I want to adjust. I asked the AI to erase the relatively prominent obstruction on the right side of the image below. As you can see, the obstruction in the lower right was completely erased, and the ground connection is very natural. There is a slight flaw in the upper right, where an unknown black object appeared.

In the Mail app, AI automatically helps categorize and puts priority emails first. In addition, a “Generate Summary” feature has been added at the top of each email.

App notifications are the same. On first use, users need to manually turn on “Generate Summary Notifications” in Settings, and can also choose which type and which app need summaries. Currently, the news and entertainment types do not yet support generating summaries. Messages sent to me by friends will appear in priority notifications.

There is another extra feature I found: apps such as Messages and Home have options for important reminders and instant notifications. This means that these apps will also display on the lock screen and play sound when Focus mode is on or the phone is muted.

The feature below is underpinned by Apple's in-house capabilities, while Qwen handles the part that "gives it superpowers."

In text input interfaces such as Messages and Mail, when I select text, the Apple Intelligence logo appears; after tapping it, the AI can help you proofread and rewrite replies. However, this rewriting cannot currently be customized—it only offers a few options: Friendly, Formal, Concise, as well as Summary, Key Points, List, and Table.

I tried using it to rewrite a reply to an unusual order inquiry. When I chose Concise mode, the generated text basically only slightly adjusted the tone and added a few emoji, with almost no change to the content; switching to Formal mode went completely off track, rewriting our reply as if it were the platform's official external messaging.

In the creation bar at the very bottom, that is where Qwen comes into play. I can give Qwen specific creative requirements, and Qwen's generation is also more complete—while filling in the request and adjusting the tone, it also generates the corresponding email format.

Visual Intelligence is equivalent to giving Siri eyes, but understanding the scene in front of it still requires Qwen to step in.

But if you want to follow up with questions or get an in-depth interpretation, you have to rely on Qwen to handle it. The "Ask" button on the lower left of the page can further invoke Qwen; the first time it is used, it will ask whether to enable it. I asked Qwen to help put together a dining suggestion for two people.

It is worth mentioning that there is also a "Search" button in the lower right corner of the Visual Intelligence page; tapping it opens Baidu Image Search and Poizon.

When encountering something it cannot solve on its own, Visual Intelligence will proactively invoke Qwen. When I opened Visual Intelligence, pointed it at a photo I had just taken in a theater, and then asked the AI to help me generate copy suitable for posting on Moments, it reminded me the first time that Qwen needed to be used for processing.

When I pointed Visual Intelligence at two posters and asked Siri what they were, Siri also invoked Qwen to help explain.

As for the Siri everyone cares about most, it can now support text input, and it can understand temporary changes in requirements during a conversation.

When I asked Siri what private home-style restaurants nearby are suitable for two people, Siri requested Maps permission and then gave the addresses of several restaurants.

When I asked it, "At 5 p.m. on Friday, will there be traffic from the China National Convention Center to Beijing South Railway Station?", Siri did not understand my specific need; it merely opened the Maps app and showed the current estimated arrival time and distance. When I asked Siri, "What's the weather like outside today, and what should I wear if I go out for a run?", Siri remained silent and just kept pushing the Weather app to me.

It can be seen that for more complex needs—such as hailing a ride for me, generating posters, or checking routes—Apple Intelligence cannot complete them directly, and users need to manually open the Qwen app for further operations.

Of course, for now only iPhone comes first. Once Mac and iPad are fully connected to Apple Intelligence later, combined with Qwen's capabilities to fill the gaps, cross-device integration may deliver a different experience.

But the AI phone battlefield in China has been fiercely competitive for several years, and real-time translation, photo erasure, AI summaries, and intelligent creation are no longer novel selling points. The current focus of domestic AI phones has shifted toward the agent direction, including support for linking complex tasks with a single sentence, proactively perceiving needs and providing services in advance, and combining long-term memory capabilities so that the system assistant can coordinate various agents to complete an entire set of work. By comparison, this China-market Apple Intelligence upgrade is still at the stage of filling in individual AI capabilities.

In the future, as devices such as Mac and iPad gradually catch up and cross-device ecosystem integration is fully connected, this combination of "Apple foundation + domestic large models" may unleash more potential.

This article comes from the WeChat public account "Zhidongxi" (ID: zhidxcom), author: Cheng Qian, published by 36Kr with authorization.