iPhone 国行 AI 真的来了，千问成为 Siri 新搭档

iPhone 国行 AI 真的来了，千问成为 Siri 新搭档

等了整整两年，中国 iPhone 用户终于不用再对着苹果 AI 的宣传片望梅止渴了。

刚刚，苹果正式宣布，Apple Intelligence（Apple 智能） 即将面向中国大陆用户推出首批功能，

从 2024 年苹果高调宣布进军生成式 AI，到海外用户陆续体验各种新功能，再到如今国行版本确定上线时间，苹果总算要兑现迟到许久的承诺。

按照计划，首批 Apple 智能功能将于 11 月 4 日随最新的 iOS 开发者 Beta 版开放测试，并于今年 12 月随 iOS 27.3 正式版面向中国用户推出。

首发设备仅限于 iPhone，包括 iPhone 15 Pro 系列、iPhone 16 及后续机型、iPhone Air 和 iPhone Duo，支持简体中文和英文。iPad 和 Mac 等其他平台的支持，则需要等待相关审批完成。

不过，此次登陆中国大陆的 Apple 智能，主要涵盖写作工具、照片编辑、智绘表情、实时翻译、通知管理、视觉智能及 Siri 基础交互升级等功能，距离大家想象中无所不能的智能助手，恐怕还有一段距离。

其中最直观的变化，是 AI 开始融入 iPhone 原本的系统功能，用户无需每次都打开独立的聊天应用。

例如，写作工具可以在系统应用及大量第三方应用中调用，支持改写、校对和摘要，只要界面中出现输入光标的位置，均可调用该工具。

给主管发送邮件时，可以选择「正式风格」优化措辞；面对备忘录里冗长的会议记录，也能直接提炼重点，整理成适合发送的工作摘要。

照片 App 则支持自然语言搜索，例如输入「毛毛在草地上」，就能找到对应的宠物照片或视频。

新增的消除工具依靠三套专用机器学习模型，分别完成物体识别、边缘判断与背景填补，整个处理过程无需上传云端。此外，用户还能根据文字描述制作回忆影片。

喜欢在聊天时斗图的用户，也有了新的玩法。苹果推出的「智绘表情」功能，可以根据简单的文字描述生成个性化表情符号，例如输入「戴墨镜的熊猫」，系统就能生成对应的表情，让用户直接在聊天中使用。

实时翻译也是此次更新中比较实用的功能。 苹果将其整合进信息、电话、FaceTime 和 AirPods 四种原生场景，支持文字和语音双向翻译，整个过程在设备端完成。

例如，前往旧金山旅行时，收到当地导游发来的英文信息，可以直接查看中文翻译，回复时也能显示双语内容。即使对方使用 Android 手机，通过 RCS 或 SMS 发送的信息同样可以翻译。对于经常联系的外国朋友，还能单独开启自动翻译。

此外，苹果还升级了通知管理功能，AI 可以将多条消息整理成摘要，通过「减少干扰」模式过滤无关提醒，并利用「优先通知」突出重要信息。邮件和信息 App 支持内容摘要与智能回复，备忘录和电话录音也能生成转写摘要。

Siri 则换上了全新的视觉设计，唤醒时屏幕边缘会出现彩色动态光效，也就是大家熟悉的「跑马灯」。同时支持语音与文字输入切换，提升了对口误、停顿和临时改口的理解能力。不过，真正能够理解个人上下文、屏幕内容并执行跨应用操作的新一代 Siri AI，仍未包含在此次更新中。

但此次升级并不包含苹果此前展示的新一代 Siri AI，也没有覆盖今年 6 月发布的最新 Apple 智能功能。 换句话说，国行用户终于能够用上苹果 AI，但那个能够理解个人上下文、屏幕内容并执行跨应用操作的 Siri，还得继续等待。

除了功能本身，国行 Apple 智能的另一大看点，是苹果终于明确了与中国 AI 厂商的合作方式。

此前，APPSO 独家获悉，苹果与阿里巴巴的合作至少涉及能力集成与联合训练两种形式。如今，首批功能已经展示出一套相当清晰的本土化技术架构。

Apple 智能的基础能力由苹果自主研发的模型提供，针对中国市场完成专项适配，首批功能全部在设备端运行。 无论写作、照片编辑还是实时翻译，个人数据都无需上传苹果服务器，苹果自身也无法访问。

但复杂任务往往需要更强大的模型，因此苹果还引入了阿里巴巴旗下的通义千问，为 Siri、写作工具和视觉智能提供能力扩展。

例如，对 Siri 说「帮我制定一份三天两夜的昆明旅行计划」，系统可以判断是否需要调用千问，在征得用户同意后完成请求，结果直接呈现在当前界面，无需跳转应用。

写作工具同样增加了千问创作选项，甚至可以将个人介绍改写成五言律诗。视觉智能则接入通义千问和百度图像搜索，帮助用户理解周围环境，并对现实世界中的文本进行翻译、朗读或总结。

值得注意的是，苹果并未绑定某一个特定版本的千问模型，而是根据任务调用多个模型，并计划持续更新。

在隐私方面，两套模型的运行机制有所区别。苹果的基础功能全部在设备端执行，千问的扩展能力则依靠服务器端模型，用户可以自行决定是否启用、分享哪些信息，并选择通过千问账户或 Apple 账户登录。

根据双方合作安排，经千问处理的用户请求数据不会用于模型训练、广告营销或行为追踪。

从产品设计来看，苹果依然坚持让自研模型负责基础功能和个人信息处理，再利用外部大模型扩展复杂任务能力。 用户不必关心背后运行的是哪套模型，只需要向 Siri 或系统工具提出需求。

苹果还明确了中国区功能的适用规则，包括大陆购买的设备，以及符合相应地区和 Apple 账户条件的 iPhone。符合条件的用户即使出国旅行，中国区 Apple 智能也能继续使用。

时间回到 2024 年 6 月的 WWDC，苹果首次发布 Apple 智能，试图让 AI 深度融入 iPhone、iPad 和 Mac，并通过更聪明的 Siri，理解个人信息、屏幕内容甚至跨应用完成复杂操作。

然而，两年过去，海外市场的基础功能虽然陆续上线，真正承担个人智能助手愿景的新版 Siri 却多次延期，中国大陆用户还经历了漫长的本土化等待。

当国产手机厂商已经将 AI 修图、实时翻译、智能摘要甚至跨应用自动操作变成常规功能，国行 iPhone 才终于迎来苹果官方的 AI 能力。

写作、翻译和照片处理固然实用，却已经算不上 AI 手机时代多么稀缺的能力。苹果真正值得期待的，是它有机会让 AI 从提供答案的工具，成为真正了解用户、能够主动完成任务的 Personal Agent。

过去两年，随着端侧 AI 快速发展，智能手机正在成为 Personal Agent 最重要的载体之一。未来的智能助手或许只需要一句话，就能结合个人习惯与实际需求，调用不同服务，完成订票、安排行程、整理工作等任务。

从这个角度看，苹果可能是最适合做 Personal Agent 的科技公司之一。

毕竟，苹果同时掌握手机、电脑、手表等硬件入口，以及操作系统和庞大的应用生态。日历、邮件、照片、健康数据与使用习惯，都有机会在用户授权下成为 AI 理解个人需求的依据。

苹果还能从操作系统层面设计 Agent 的工具调用与跨应用协作机制。此次展示的端侧模型与千问协同架构，也让人看到了简单任务本地处理、复杂任务调用外部模型的可能性。

但有意思的地方在于，苹果可能恰恰又是最不适合做 Personal Agent 的公司之一。

苹果长期强调隐私保护、应用隔离与用户控制权，而真正聪明的 Agent，需要关联不同应用的信息，甚至主动执行操作。当 AI 开始读取邮件、查询日历乃至代替用户支付和预订，数据权限与误操作风险都会变得更加复杂。

更深层的矛盾来自苹果自己建立的 App 生态。当 AI 能够替用户购物、订酒店、安排出行，用户可能越来越不需要亲自打开 App。届时，苹果经营十几年的 App Store 应用分发与商业生态，也可能因此受到挑战。

对于一家长期依靠软硬件协同、严格管理用户体验，并从应用生态中获得丰厚收益的公司而言，真正自主的 AI Agent，意味着它需要重新思考操作系统、应用与用户之间的关系。

回到此次国行 Apple 智能的发布，苹果总算让中国用户拥有了体验官方 AI 的机会，也展示了自研端侧模型与本土大模型协同工作的能力。

但相比已经登场的写作工具、实时翻译和智绘表情，我更期待未来的 Siri 能够真正理解个人上下文，并在用户授权下完成跨应用操作。

毕竟，苹果已经具备创造 Personal Agent 的许多先天条件。手机这个行业，上一次被彻底改变，是 iPhone 定义的触屏时代。而下一次改变，大概率来自 AI。

苹果亲手开启了上一个时代，也有很大的机会开启下一个。只是这一次，它要超越的对手不再是诺基亚，而是那个曾经敢把一切推倒重来的自己。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

IPhone National Bank AI is really here.

After two years of waiting, China's iPhone users finally stopped thirsty for Apple AI's promotional film.

Just now, Apple has officially announced that Apple Intelligence is about to launch its first features for mainland Chinese users.

From Apple's high-profile march-generation AI in 2024, to its overseas users experiencing new features, to its current state-of-the-art version, which determines when to go online, the Apple will finally meet its late promise.

It is planned that the first Apple smart features will be launched to Chinese users on November 4 with the latest iOS developer, Beta, and the official iOS 27.3 in December this year.

Launch equipment is limited to iPhone, including iPhone 15 Pro series, iPhone 16 and follow-up, iPhone Air and iPhone Duo, in support of Chinese and English. Support from other platforms, such as iPad and Mac, will need to await approval.

However, the Apple intelligence that landed in mainland China, which mainly covers writing tools, photo editing, intellectual expression, real-time translation, notification management, visual intelligence, and the Siri Foundation Interactive Upgrading, is still a distance from what you can imagine as an incompetent smart assistant.

The most intuitive change is that AI begins to integrate the iPhone original system functionality, and users do not have to open an independent chat application every time.

For example, the writing tool can be called in system applications and in a large number of third-party applications, supporting rewrite, proofreading and summary, and can be called whenever the location of the input cursor appears in the interface.

When sending an e-mail to the supervisor, you can opt for the optimal wording of "official style" and, in the face of lengthy minutes in the memo, you can extract the focus directly into a job summary suitable for dispatch.

Photo App supports the search for natural languages, for example, by entering "the fur on the lawn" and finding a matching pet photograph or video.

The new elimination tool relies on three specialized machine learning models, each completing object identification, edge judgement and background filling, with no need to upload clouds throughout the process. In addition, users can produce memoirs based on the description of the text.

There are new ways to play people who like to talk to each other. Apple's introduction of the "Put-in-the-Wise Emoticon" function allows for the generation of personalized emoticons based on simple text descriptions, such as the input of the "Panda with sunglasses", and the system produces the corresponding expression, allowing users to use it directly in the chat.

Real-time translation is also a useful feature in this update. Apple integrates it into four original scenarios of information, telephone, FaceTime and AirPods, supporting translation of text and voice in two directions, and the process is completed at the end of the device.

For example, when travelling to San Francisco, information received from local guides in English is available for direct viewing of the Chinese translation and bilingual content when responding. Even information sent through RCS or SMS can be translated using Android phones. Automatic translations can also be opened separately for foreign friends who are in constant contact.

In addition, Apple has upgraded the notification management function, which allows AI to organize multiple messages into summary, filter unrelated alarms through the "reduce interference" mode and use "prior notice" to highlight important messages. Mail and Information App supports content summary and smart responses, and memos and phone recordings can also generate summary.

Siri was replaced with a brand new visual design, with color dynamic light at the edge of the screen, which is known as the "run horselight". Simultaneously, voice-to-text switching is supported, increasing the ability to understand verbal errors, pauses and temporary re-transformations. However, a new generation of Siri AI that can truly understand the individual context, screen content and cross-application is still not included in this update.

But this upgrade does not include the new generation of Siri AI that Apple had previously displayed, nor does it cover the latest Apple smart feature of June this year. In other words, the national user can finally use Apple AI, but Siri, who understands the personal context, screen content and cross-applications, has to wait.

Apart from the function itself, another big point of view of Apple's intelligence is that Apple finally clarified the way in which it works with Chinese AI manufacturers.

Previously, APPSO alone had learned that Apple ' s cooperation with Alibaba involved at least two forms of capacity integration and joint training. Today, the first functions have demonstrated a fairly clear indigenous technological architecture.

Apple ' s basic capabilities are provided by a model developed by Apple on its own, which is tailored to the Chinese market, with all the first functions running at the end of the equipment. Neither writing nor photo editing nor translation in real time, personal data need to be uploaded on the Apple server, and apples cannot be accessed by themselves.

However, complex tasks often require stronger models, so apples also introduce Qwen under the Alibaba flag, providing capacity expansion for Siri, writing tools and visual intelligence.

For example, in the case of Siri, who said "help me develop a three-day, two-night Kunming travel plan", the system can judge whether it is necessary to call and complete the request with the user's consent, and the result is presented directly at the current interface and does not need to jump-start the application.

Writing tools have also added creative options and can even be translated into five poems. Visual intelligence is linked to Qwen and Baidu image searches, helping users understand the surrounding environment and translating, reading or summing up texts in the real world.

It is worth noting that apples do not bind a particular version of the question-and-answer model, but rather call multiple models according to the mission and plan to update them on a continuous basis.

In terms of privacy, the operating mechanisms of the two sets of models differ. The basic function of apples is fully performed at the end of the device, while the extension capability depends on the server-end model, which allows users to decide whether to enable, share and choose to log in through the account or Apple.

Under cooperative arrangements, user requests for data processed after a thousand inquiries will not be used for model training, advertising or behavioural tracking.

In terms of product design, apples continue to insist that self-study models are responsible for basic functions and personal information processing, and that external LLM be used to expand complex mission capabilities. Users need not care about which models are running behind them, but simply ask for Siri or system tools.

Apple also specified the applicable rules for the functioning of the Chinese district, including equipment purchased by the mainland and iPhone that met the conditions of the respective area and Apple account. Even if the eligible users travel abroad, Apple intelligence in the China region can continue to be used.

Back in June 2024, WWDC, Apple's first publish Apple intelligence, tried to get AI deep into iPhone, iPad and Mac, and through a smarter Siri, to understand personal information, screen content and even cross-application.

However, two years later, while the basic functions of overseas markets have been on the line, the new version of Siri, which truly embraces the vision of personal intelligence assistants, has been postponed several times, and users in mainland China have been waiting for long periods of localization.

When the national mobile phone manufacturer has turned AI fixes, real-time translations, smart summarys, and even cross-application automatics, the country bank iPhone finally reaches the official Apple AI capability.

Writing, translation and photo processing, while useful, are no longer as scarce as the capacity of the AI mobile phone age. What Apple really deserves is that it has the opportunity to give AI the tools to provide answers to become a person who really knows the users and can do the job.

In the last two years, as the end side of AI has developed rapidly, smartphones are becoming one of the most important carriers of personal age. Future smart assistants may need only one sentence to combine personal habits with actual needs, to call on different services, to order tickets, schedule trips, organize work, etc.

From this point of view, apples may be one of the most suitable technology companies for being personal Agent.

After all, apples also have access to hardware such as mobile phones, computers and watches, as well as operating systems and vast application ecology. Calendars, mail, photographs, health data and usage habits have the opportunity to be the basis of AI ' s understanding of individual needs with user authorization.

Apple can also design Agent ' s tool call and cross-application collaboration mechanisms at the operational system level. The end-side model presented this time, in conjunction with the multi-question architecture, also showed the possibility of simple tasks being handled locally, complex tasks calling on external models.

But what's interesting is that apples are probably one of the most unsuited companies in the world.

Apple has long emphasized privacy protection, application isolation and user control, while truly intelligent Agent needs information linking different applications and even proactive operations. Data permission and error risks become more complex when AI starts reading mail, searching calendars and even replacing user payments and reservations.

Deeper contradictions come from Apple's own App Ecology. When AI is able to shop for users, book hotels, arrange trips, users may be less and less obliged to open App. At that time, Apple's more than a decade-old app distribution and commercial ecology could also be challenged.

For a company that has long relied on the synergy of software and hardware, that strictly manages the user experience, and that reaps significant benefits from the application of ecology, a truly autonomous AI Agent means that it needs to rethink the relationship between operating systems, applications and users.

Back to Apple’s smart publish, Apple finally gave Chinese users the opportunity to experience official AI and demonstrated the ability of self-study side models to work with home-grown LLM.

But I'm looking forward to the future Siri actually understanding the individual context and completing the cross-application with the user's permission, compared to the writing tools, real-time translation and intellectual expression that have already taken place.

After all, apples already have many congenital conditions for creating Personal Agent. The cell phone industry, last time completely changed, was the iPhone-defined touch screen era. And the next change is probably from AI.

Apple itself opened the last era, and there was a great opportunity to start the next one. Only this time, it is no longer Nokia, but the one who dared to push everything back.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.