实测 iPhone 国行 AI:有亮点有槽点,还有一个惊喜点Testing iPhone's China-edition AI: highlights, gripes, and one pleasant surprise


AI free token/AI免费积分

就在刚刚,随着 iOS 27.3 开发者测试版更新,笔者苦等两年多的苹果智能,终于向国内使用者开放测试了!

从 2024 年 WWDC 首次亮相,到如今真正装进国行设备里,苹果这张画了两年的 AI 大饼,总算端到了笔者面前。

更关键的是,苹果这次给国行版找了个不错的本土搭档:千问。

这次的国行苹果智能,由苹果自研模型与阿里千问共同撑起底座。诸如照片消除、基础润色这类能力直接在端侧运行;而一旦遇到复杂任务,平台则会借助 Siri、写作工具和视觉智能去调用千问的云端能力。

Image

至于哪些 iPhone 能用、国行与海外版有何区别,以及 Apple 账号和地区限制等难题,笔者此前已经做过详细梳理,感兴趣的朋友可以移步阅读刚刚,国行 Apple 智能官宣,但满血 SiriAI 还得等。

等了这么久,国行 Apple 智能到底实际表现怎么样?

在正式看感受之前,如果你手里正好有受兼容的机型,也想第一时间尝鲜国行 Apple 智能,可以借助以下步骤获取这次的开发者测试版:

先在 App Store 搜索并下载官方的「Apple Developer」应用,在账户栏登录自己的 Apple ID;稍等片刻后,打开平台的「设置 – 通用 – 软件更新」,就能在「Beta 版更新」里刷出最新的 iOS 27 Developer Beta 了。

毕竟是首版开发者测试固件,平台里不可避免地存在一些 Bug ,如果你手里只有一台日常高频使用的主力机,建议等到 12 月正式版发布后再稳妥上车。

Siri、修图、翻译好用吗?

本次平台更新的包体巨大,足足有 12GB,升级前请务必为手机留出足够大的存储空间。

且由于国行 AI 大量能力依赖端侧模型运行,需要建立新的资料检索,因而在刚上手的当下,存在一些稳定性难题,譬如相册里的自然语言搜索、回忆影片,以及锁屏的通知总结等能力都需要等待索引完成,可能需要等待一到两天才能能正常使用。

接下来,挑几个重点能力为人们实测一下:

全新 Siri

打头阵的,当然是 Siri。告别了那个只会设闹钟、查天气,稍微问深一点就甩给你几个网页链接的「人工智障」,换上跑马灯新皮肤的 Siri,终于兼容双击显示屏底部横条,直接打字交互了。

Image

实测下来,如果你只是问点简单的日常难题,或者直白地下达基础设置指令,它能直接搞定。但只要你稍微绕个弯,Siri 仍然是会蒙圈。

至于一句话塞两个任务、或者更深层的复杂平台调度,它仍旧无能为力——就平台控制力而言,它骨子里还是以前那个人们熟悉的 Siri。

那遇到更复杂的长难任务呢?笔者直接丢给它一个写攻略的需要:「我下个月想去九寨沟玩四天,帮我安排一份旅游攻略。」

Image

面对这种超出自身能力范围的难题,国行版 Siri 会弹窗询问你是否同意调用千问,并提示需要登录千问账号(如果你在这里挑选拒绝调用,Siri 会干脆利落地直接放弃回答。)

在点击同意调用之后,大概 7、8 秒之后就会生成答案。单看生成速度确实不慢,但细看内容就有点露怯了。它给出的所谓四日攻略,不论是住宿建议还是景点安排都写得极其简略,信息密度很低,顶多只能算是一个格外初步的大纲预览,实际成效是远不如在千问 app 里直接提问的。

Image

图片消除

相比仍旧有些木讷的 Siri,照片 App 里的「消除」能力感知要直接得多,这项能力调用的是苹果的端侧模型。

实测下来,苹果智能的消除能力表现还是挺可圈可点的。

具体来说,如果画面里要抹掉的主体离镜头比较远、轮廓清晰,而且周围背景比较干净,消除的成效格外干净利落,而且画质损失很小,单张照片的体积大概只会减少 0.2MB 左右。

Image

可一旦你要求删除的主体比较复杂——例如笔者测试图中大面积的草地和鹅卵石路面,就比较容易翻车。消除完之后,被涂抹的区域会留下一块分辨率极低的模糊色块,跟周围原本清晰的画质格格不入。

Image

实时翻译

另一个感受下来能打好评的能力,是 Apple 智能的「实时翻译」。

这项能力同样依赖于端侧模型,初次使用时需要下载对应翻译语言的语言包,每个大概在 2GB 左右。

苹果把这项能力塞进了信息、电话、FaceTime 等应用里,甚至连接 AirPods 时,还能达成「边聊边译」。

实测下来,实时翻译能力在日常口语对话时表现让人惊喜,它的反应速度很快,延迟不到 1 秒,准确率也还不错,足够支撑你和老外顺畅交流。

Image

写作工具

在任意可以开展文本输入的地方,选中文字,就能调用写作工具开展改写、校对、总结,写作工具绝对是打工人最容易用上的能力。但对于国行版本,笔者更关心的是,这套从海外引进的 AI 写作工具,会不会在中文语境里「水土不服」?

例如一段比较生硬的催稿消息:「你们之前说周五交思路,现在已经周一了,为什么还没发过来?请尽快处理。」用写作工具尝试改写,哪怕笔者特意切到了最温和的「友好」模式,改出来的句子还是会有一板一眼的感觉。以及涵盖更多的摘要、要点等能力,都跟大部分的 ChatBot 的水平类似。

Image

不过在调用的便捷度上,苹果倒是发挥了平台级的长处。实测只要是有文字输入框的地方——不管是平台自带的备忘录,还是微信、飞书、钉钉这类第三方应用,都能从容的去调用,还是比较方便的。

Image

视觉智能

苹果这次还把识别能力搬到了现实世界——视觉智能,长按「相机控制」按键就能激活,主要有两种交互入口:「询问」和「搜索」。

左边的「询问」接入了阿里千问的能力,本意是让你直接针对眼前的物品提问,例如拍一下护肤品,顺便问问它的核心成分和用途。

Image

右边的「搜索」则接近笔者熟悉的「以图搜图」,默认使用的是百度的视觉模型,可以快速检索到百度上能搜到的同类图片。

但笔者在测试时发现了一个惊喜点:除了百度识图以外,许多 app 借助接入对应的 API 接口,也能给出类似的能力。

譬如在手机里安装了得物,它就浮现在了可选的搜索服务列表中,并检索到得物里在售的类似物品;甚至连我手机里装的一款冷门和弦查询应用 Guitar Wiz,也浮现在了可选的搜索服务列表中,可以用于查询和弦。

不难推测,苹果大概率为第三方应用开放了对应的 API 接口——只要开发者做过适配,使用者手机里装了对应的 App,就能直接借助视觉智能调起它的搜索能力。笔者很期待淘宝、小红书等应用接入后的表现。

除了以上 AI 能力之外,国行 AI 还兼容智慧表情(Genmoji)能力,在聊天斗图的时候,可以方便地做点表情包:

Image

幽默。

苹果离真正的智能,还有多远

把国行 Apple 智能的首批能力挨个摸了一遍后,平时关注苹果的朋友心里应该都有数——说白了,这就是 2024 年海外最早推送的「初版」 Apple Intelligence,并非苹果最新的手机 AI 应对思路。

Image

事实上,在 iOS 27.3 的海外测试版中,SiriAI 已经兼容繁体中文,并且能力上有了进一步的升级,愈发有成熟的趋势。

比较之下,刚刚更新的国行苹果智能多少显得有些不够「智能」。

但,这仍旧是里程碑式的进步。

苦等了两年多,中国使用者的 iPhone 总算是把这块缺失已久的平台级 AI 拼图给补上了。更何况对绝大多数普通人来说,人们本来也没指望手机里突然蹦出个无所不能的科幻管家,无非是希望 Siri 别再像以前那样一问三不知,能用上像实时翻译、图片消除等好用的能力。

倘若细细追究,这台 iPhone 有没有因为 Apple 智能而发生什么「质变」?

显然没有。

当下的这批能力,本质上还是「AI +」的赋能思路——相册加上 AI,变成了照片消除;输入框加上 AI,变成了文字润色。在这里,AI 只是给原有应用打辅助的「外挂插件」,而不是手机的底层逻辑。

Image

笔者真正期待的,是 AI 能接管底层交互:当你对手机说一句「把朋友刚发的餐厅地址加进今晚日程,再看看从公司过去要多久」时,它能认出朋友、找到地址、写入日程,替你从头到尾把事情办妥。

实际上苹果今年推出的新版 Siri AI,就是主打显示屏感知和情境理解,并具备一定的个人上下文管理能力。

即便如此,过去这段时间绝大多数使用者,对 Apple Intelligence 乃至新版 Siri AI 的反馈也相当冷淡:

感受远没有宣传片里那么神。

爱范儿认为,造成这种落差的关键,还是在于长程行动能力与个人记忆管理。

手机,或许是每个人最亲近的电子设备,不仅能在后台长期记住你的生活轨迹——分辨出哪个是你的「朋友」,找到你的「公司」在哪,还能跨越权限去翻找聊天记录、调动日历和地图。

当下来看,苹果正在努力打通 Siri 与 iPhone 上的各个原生应用,使其数据得以流转,调度更加顺畅。

Image

但笔者的数字生活还高度绑在微信、美团、飞书这些三方 App 里,苹果就算把自家的备忘录和日历打通得再丝滑,只要进不去第三方应用的围墙,它对你的理解就是断层的。

平心而论,即便笔者很快就能顺利用上国行版的新 Siri AI,短期内对笔者的日常生活也不会有翻天覆地的影响。对于苹果而言,如何建立好个人上下文,让手机里的 AI 可以更好服务于人,仍旧是一件需要很长时间去沉淀的事情。

过去两年,笔者很多发布会上看过太多眼花缭乱的「一句话点咖啡」演示,但真正把手机当主力机用上几个月就会发现,演示 Demo 里的惊艳,离每天日常使用的稳定可靠,中间还隔着十万八千里。

个人 AI 仍旧处于投入期,可却也是被所有人寄予厚望的「第二增长曲线」。

还是以苹果为例,尽管刚刚交出有史以来最好的财年财报,但仍面临巨大的挑战——

硬件利润正在被元器件涨价潮侵蚀,应用分成又被全球监管盯着不放,这时苹果的 AI 战略,既是守住硬件基本盘的防御塔,也是服务业务最顺理成章的增长点。

面对手握超过 25 亿台活跃设备、15 亿份付费订阅的庞大基本盘,苹果 AI 的商业价值不只是吸引消费者更新设备,也是使用者多掏钱的理由。

在这次财报电话会上,苹果高管几乎把这套商业阳谋摆在了台面上。CFO Kevan Parekh 明确提到,AI 将借助自有服务与合作伙伴创造新的变现机会;新任 CEO John Ternus 更是直接剧透,将来像 Siri AI、高阶智能修图这类依赖云端大模型的能力,将会设置基础使用额度,重度使用者需要付费购买更多额度的服务。

Image

这也意味着,过去主要云端存储空间来吸引使用者掏钱的 iCloud+,将来很可能会摇身一变,成为苹果兜售「云端算力与高阶智能」的订阅入口。苹果显然不打算只凭这一代机型的表现,来衡量这场长期投资的价值。

只不过,商业蓝图再性感,那也是苹果给华尔街讲的远期故事;回到笔者自己手里这台 iPhone 上,眼下的预期还是不妨放平一些。

国行 Apple 智能的落地,说到底还是一次迟到的补课——它总算把中国使用者缺失的那些 AI 基础能力给补上了。

至于什么时候才能用上那台真正的 AI Phone,无论苹果还是手机行业,或许还得再之类。

本文来源:https://www.ifanr.com/1687733
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

Just moments ago, with the release of the iOS 27.3 Developer Beta update, the Apple Intelligence feature that I have eagerly awaited for over two years is finally open for testing to users in China!

From its first appearance at WWDC 2024 to being truly installed into devices in mainland China today, this AI pie that Apple has been drawing for two years has finally been served to me.

More crucially, Apple has found a solid local partner for the mainland China version this time: Qwen.

This version of Apple Intelligence for mainland China is built on a foundation jointly supported by Apple's in-house models and Alibaba's Qwen. Capabilities such as Photo Erase and basic text polishing run directly on-device; meanwhile, for complex tasks, the platform leverages Siri, Writing Tools, and Visual Intelligence to invoke Qwen's cloud capabilities.

Image

Regarding which iPhones are compatible, the differences between the mainland China and overseas versions, and issues such as Apple Account and regional restrictions, I have previously provided a detailed breakdown. Interested friends can check out the article "Just Now, Mainland Apple Intelligence Officially Announced, but Full-Blooded Siri AI Still Needs to Wait."

Having waited this long, how does the actual performance of Apple Intelligence in mainland China hold up?

Before diving into the review, if you happen to have a compatible device and want to be among the first to try out Apple Intelligence in mainland China, you can follow these steps to get the Developer Beta:

First, search for and download the official "Apple Developer" app on the App Store, and sign in to your Apple ID under the Account section; after a moment, open the device's "Settings – General – Software Update," and you will see the latest iOS 27 Developer Beta available under "Beta Updates."

Since this is the first Developer Beta firmware, it inevitably contains some bugs. If you only have one main device for daily high-frequency use, it is recommended to wait for the official release in December to make the switch safely.

Is Siri, photo editing, and translation any good?

The update package for this platform update is massive,足足有 12GB, so please make sure to leave enough storage space on your phone before upgrading.

Also, because many AI capabilities for mainland China rely on on-device models, a new data index needs to be established. Therefore, right after the initial setup, there are some stability issues; features such as natural language search in Photos, Memories movies, and lock screen notification summaries require waiting for the indexing to complete, which may take one to two days to function normally.

Next, I will test a few key capabilities for you:

The All-New Siri

Leading the charge is, of course, Siri. Gone is the "artificial stupidity" that could only set alarms, check the weather, and throw a few web links at you when asked anything deeper. Donning a new "running light" skin, Siri finally supports double-tapping the horizontal bar at the bottom of the display to interact via direct text input.

Image

In testing, if you just ask simple daily questions or give straightforward basic command settings, it can handle them directly. But as soon as you beat around the bush even slightly, Siri still gets confused.

As for fitting two tasks into one sentence, or deeper complex platform scheduling, it remains powerless—in terms of platform control, it is still the same familiar Siri in its bones.

So what about more complex, lengthy tasks? I threw a request for a travel guide at it: "I want to visit Jiuzhaigou next month for four days; help me arrange a travel itinerary."

Image

Faced with a question beyond its capabilities, the mainland China version of Siri pops up asking if you agree to invoke Qwen and prompts you to log in to a Qwen account (if you choose to refuse the invocation here, Siri will simply give up answering immediately).

After clicking agree, the answer is generated in about 7 or 8 seconds. The generation speed is indeed fast, but looking closely at the content, it reveals some weakness. The so-called four-day guide it provided is extremely brief regarding accommodation suggestions and attraction arrangements, with very low information density. It is at best a very preliminary outline preview, and the actual effectiveness is far inferior to asking directly in the Qwen app.

Image

Photo Erase

Compared to the still somewhat wooden Siri, the perception of the "Erase" capability in the Photos app is much more direct; this capability utilizes Apple's on-device model.

In testing, the performance of Apple Intelligence's Erase capability is quite commendable.

Specifically, if the subject to be removed is far from the lens, has clear outlines, and the surrounding background is relatively clean, the erasure is exceptionally clean and crisp, with minimal image quality loss; the size of a single photo decreases by only about 0.2MB.

Image

However, once you ask to delete a more complex subject—such as the large area of grass and cobblestone paths in my test image—it is prone to failure. After erasure, the smeared area leaves behind a low-resolution blurry color block that clashes with the originally clear image quality of the surroundings.

Image

Real-Time Translation

Another feature that deserves high praise after hands-on experience is Apple Intelligence's "Real-Time Translation."

This capability also relies on on-device models. When using it for the first time, you need to download language packs for the corresponding translation languages, each approximately 2GB in size.

Apple has integrated this capability into apps like Messages, Phone, and FaceTime. Even when connected to AirPods, it enables "translate while you chat."

In real-world testing, the real-time translation capability performed surprisingly well in everyday spoken conversations. Its response speed is fast, with latency under 1 second, and accuracy is quite good, sufficient to support smooth communication with foreigners.

Image

Writing Tools

Anywhere you can input text, simply select the text to call up Writing Tools for rewriting, proofreading, and summarizing. Writing Tools is definitely the feature that office workers will find most useful. However, for the China mainland version, what the author is more concerned about is whether this AI writing tool imported from overseas will experience "acclimatization issues" in the Chinese context.

For example, take a rather blunt message urging a draft: "You said you'd submit the outline on Friday, and now it's already Monday. Why hasn't it been sent yet? Please handle this ASAP." When trying to rewrite it with Writing Tools, even though the author specifically switched to the mildest "Friendly" mode, the rewritten sentences still felt stiff and formulaic. As for other capabilities such as summaries and key points, they are similar to the level of most ChatBots.

Image

However, in terms of accessibility of invocation, Apple has leveraged its platform-level strengths. In testing, anywhere there is a text input field—whether it's the platform's built-in Notes app, or third-party apps like WeChat, Feishu, or DingTalk—the tool can be called up smoothly, which is quite convenient.

Image

Visual Intelligence

Apple has also brought recognition capabilities into the real world this time—Visual Intelligence. It can be activated by long-pressing the "Camera Control" button, with two main interaction entry points: "Ask" and "Search."

The "Ask" option on the left integrates Alibaba Qianwen's capabilities, designed to let you directly ask questions about items in front of you. For example, take a photo of a skincare product and ask about its core ingredients and usage.

Image

The "Search" option on the right is similar to the "image search" feature the author is familiar with. It defaults to using Baidu's visual model and can quickly retrieve similar images available on Baidu.

But during testing, the author discovered a pleasant surprise: besides Baidu image recognition, many apps can also provide similar capabilities through corresponding API integrations.

For instance, if Dewu (Poizon) is installed on the phone, it appears in the list of selectable search services and can retrieve similar items listed for sale on Dewu. Even a niche chord lookup app called Guitar Wiz installed on my phone also appeared in the selectable search services list and can be used to look up chords.

It's not hard to speculate that Apple has most likely opened up corresponding API interfaces for third-party apps—as long as developers have made adaptations and users have the corresponding apps installed on their phones, they can directly leverage Visual Intelligence to invoke their search capabilities. The author is looking forward to seeing how apps like Taobao and Xiaohongshu perform once integrated.

In addition to the AI capabilities mentioned above, the China mainland AI also supports the Genmoji feature, making it easy to create custom emojis during chats and emoji battles:

Image

Humor.

How Far Is Apple from True Intelligence

After going through all the first batch of capabilities of the China mainland Apple Intelligence, those who follow Apple regularly probably already have a good idea—put simply, this is the "initial version" of Apple Intelligence first rolled out overseas in 2024, not Apple's latest approach to mobile AI.

Image

In fact, in the overseas beta version of iOS 27.3, SiriAI already supports Traditional Chinese and has undergone further upgrades in capabilities, showing a growing trend toward maturity.

In comparison, the just-updated China mainland Apple Intelligence appears somewhat lacking in "intelligence."

But this is still a milestone achievement.

After more than two years of waiting, Chinese iPhone users finally have the missing platform-level AI puzzle piece in place. Moreover, most ordinary people never expected a phone to suddenly produce an omnipotent sci-fi housekeeper; they simply hoped Siri would stop being completely clueless as it once was, and that they could use handy features like real-time translation and photo erasure.

But if we look closely, has this iPhone undergone any "qualitative change" because of Apple Intelligence?

Clearly not.

At their core, the current batch of features still relies on the empowerment approach of "AI plus"—the photo album plus AI becomes photo erasure; the input box plus AI becomes text polishing. Here, AI is merely an auxiliary "plug-in add-on" for existing apps, rather than the underlying logic of the phone.

Image

What I truly expect is for AI to take over the underlying interactions: when you tell your phone, "Add the restaurant address my friend just sent to tonight's schedule, and check how long it takes to get there from the office," it should be able to recognize the friend, find the address, add it to the schedule, and handle the whole task from start to finish for you.

In fact, the new version of Siri AI launched by Apple this year focuses on screen awareness and contextual understanding, and possesses a certain degree of personal context management capability.

Even so, most users have responded rather lukewarmly to Apple Intelligence and even the new version of Siri AI over the past period:

They feel nowhere near as miraculous as shown in the promotional videos.

Ifanr believes that the key to this gap still lies in long-range action capabilities and personal memory management.

A phone is perhaps the closest electronic device to each person, capable not only of remembering one's life trajectory in the background over the long term—distinguishing who your "friends" are and finding out where your "office" is—but also of reaching across permissions to search chat records and mobilize calendar and map data.

As it stands, Apple is working to connect Siri with various native apps on the iPhone, enabling data to flow between them and making scheduling more seamless.

Image

However, my digital life remains highly tied to third-party apps like WeChat, Meituan, and Feishu. Even if Apple makes its own Notes and Calendar integration as smooth as silk, as long as it cannot break into the walled gardens of third-party apps, its understanding of you will remain incomplete.

To be fair, even though I will soon be able to use the China mainland version of the new Siri AI smoothly, it will not have a drastic impact on my daily life in the short term. For Apple, how to build a good personal context so that AI on the phone can better serve users is still something that will take a long time to mature and develop.

Over the past two years, I have seen too many dazzling "order coffee with one command" demos at product launch events. However, after actually using a phone as your primary device for a few months, you will find that the awe-inspiring moments in demos are still a world away from the stable reliability required for daily use.

Personal AI is still in the investment phase, yet it is also the "second growth curve" with high expectations placed on it by everyone.

Take Apple as an example. Even though it just delivered its best fiscal year financial report ever, it still faces enormous challenges—

Hardware profits are being eroded by a wave of rising component prices, and application revenue sharing is under constant scrutiny from global regulators. In this context, Apple's AI strategy serves as both a defensive tower to safeguard its core hardware business and the most logical growth point for its services business.

With a massive foundation of over 2.5 billion active devices and 1.5 billion paid subscriptions, the commercial value of Apple's AI lies not only in attracting consumers to upgrade their devices but also in providing a compelling reason for users to spend more money.

At this earnings call, Apple executives practically laid this commercial playbook out on the table. CFO Kevan Parekh explicitly mentioned that AI will create new monetization opportunities through Apple's own services and partnerships; newly appointed CEO John Ternus even directly revealed that future cloud-based large model-dependent features, such as Siri AI and advanced intelligent photo retouching, will have a basic usage quota, with heavy users required to pay for additional quota services.

Image

This also means that iCloud+, which has traditionally attracted users to pay for expanded cloud storage space, may very well transform in the future into a subscription gateway for Apple to sell "cloud computing power and advanced intelligence." Apple clearly does not intend to measure the value of this long-term investment based solely on the performance of this generation of devices.

However, however seductive the commercial blueprint may be, it is merely a long-term story Apple is telling Wall Street; when it comes to the iPhone in my own hands right now, it is still wise to temper our expectations.

The rollout of Apple Intelligence in China mainland is, in essence, just a delayed catch-up—it has finally filled in the basic AI capabilities that Chinese users had been missing.

As for when we will actually get to use a truly authentic AI Phone, both Apple and the mobile phone industry might still need to wait a bit longer.

暂无评论,快来发表第一条评论吧!No comments yet. Be the first!

📮 需求咨询📮 Inquiry