首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活Hands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac Studio


AI free token/AI免费积分

21499 元的 iPhone Duo,加上 86749 元的 Mac Studio,两台设备加起来超过 10 万元。

APPSO 想用这两台苹果史上最贵的手机和电脑,构建一个这个时代最优雅的「智能个人中心」。

Image

256G 内存,1.2TB/s 内存带宽的超强算力 Mac Studio 留在家里,负责运行模型和执行 Agent 任务。

出门只带一台 Duo,需要的时候掏出手机,就能看看家里的电脑忙到哪了,或者再给它派点活。

Image

为此笔者还专门做了一个深度适配 iPhone Duo 的 App,针对合盖、半折悬停以及完全展开三种不同的形态,分别外观了对应 Agent 的不同工厂情形。

Image

iPhone Duo + Mac Studio,苹果 AI 王炸组合

先把 Mac Studio 这台本地算力安排好,笔者就开始着手怎么让手机能连接到电脑。就像 Codex 一样,可以在手机上就下达任务,利用本地电脑的环境,执行各种复杂任务。

Image

从流程来看,主要应对的难题就是手机怎样找到 Mac、任务怎样交给 Mac、执行情况怎样回到手机。

为了让手机连进来,Mac 上另外运行一个 HTTPS 代理,代理监听 Mac 的局域网地址,收到手机请求后,转发给本机的 Bridge。

如果以后要在外面用蜂窝网络连接,就可以让手机和 Mac 加入同一个 Tailscale 私有网络,再借助 Tailscale Serve 把 Bridge 给出为私有 HTTPS 服务。

Image

连接后,App 会先检查服务是否正常,再读取会话列表和所选会话的内容。笔者选中的会话有一个 sessionId,后续任务就发送到这个会话。

任务进入 Harness 后,由 Mac 上的 Agent 调用本地模型、读取工作区、执行工具,并保存执行记录。手机连接中断不会触发任务取消;只要 Mac 和相关服务继续运行,任务仍由 Harness 管理。

首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活

合上盖,就是 Agent 仪表盘

打开 App,正面是一排任务卡片,笔者能看到哪个 Agent 在跑、跑到哪了、还要多久,都是扫一眼就完事。顶部则是 Mac Studio 的负载摘要,有内存占用、模型名称、Token 用量、运行中任务数、今日完成数等信息。

首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活

既然家里摆着一台这么贵的电脑,总得知道它现在忙成什么样。

在这个形态下最高频的操作大概是审批。如果是云端的沙箱环境,可以毫无顾忌地直接把项目权限从「手动批准」修改为「替我批准」,但如果是自己的电脑,还是得保险起见,让常规操作自动执行,关键决定再询问。

所以 Agent 干活干到一半,经常就有需要人确认的情况,在合盖模式直接把「确认 / 拒绝」两个按钮做得又大又显著,正好在拇指底下,何时何地从口袋里掏出手机打开 App 就能批准。

首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活

而针对现在 Agent 越来越长的任务,还可以挂到灵动岛上,锁屏也能看到项目进度执行到了哪里。几乎 90% 的远程交互,在这个形态里就闭环了。

半折状态是移动指挥中心

这是整台机器最「只有 Duo 能做到」的形态。

Image

把 Duo 折到一半放在桌上,它自动切成上下分工。

Image

上半屏是直播,Agent 的执行进度会像视频监控一样,自动滚动完整的轨迹,涵盖工具调用,模型的推理之类。

下半屏则是操作,触控板、终端输入框、审批按钮,还有一个语音派单键。半折状态放桌上,恰好是语音输入最自然的姿势。

首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活

实测这个形态最惬意的情形,就像是追剧一样,把 Duo 立在桌角,看着家里的机器一格一格把片子渲完。

展开是一套完整的控制台

完全展开的 Duo 则接近一台小平板,所有「在手机上太紧凑、又懒得掏电脑」的操作,全部归到这个形态。

Image

布局是像电脑上经典双栏一样,左边是任务队列,右边选中任务的完整会话。Agent 生成的产物,涵盖渲染预览图、整理好的表格、剪辑好的视频,以及网页之类,在这里有足够大的空间预览,不满意也可以更方便的批注,让它重做。

Image

Image

更有意思的是分屏打开多个 App 的操作,就像笔者刷相册看到一张参考图,可以直接拖进任务卡片,就是新任务的输入。

其次,两个 Agent 会话可以并排表明,左边看网页的生成进度,右边改另一个任务的提示词。

Image

三个形态反而成了三个 App 一样,合盖给结论,悬停给画面,展开给细节。

笔者在合盖状态看到了审批请求,展开来看详情,批准完再折回去塞回口袋。信息密度可以跟着显示屏面积走,这是折叠屏客户端和普通手机 App 最本质的区别。

基于 iPhone Duo 的几种形态,笔者还 vibe coding 了另外几个有意思的 App,让折叠这种形态发挥它真正的价值。

就像这个摸鱼的最佳姿势,外人看以为是捧着一本深度报告在读,实际上折叠的大屏部分在网上冲浪。

首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活

笔者甚至可以自定义封面,设置为高等数学等不同的课本封面。

首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活

还有依据 iPhone Duo 的形态外观对应的游戏,显示屏一合上,就是把满屏的蟑螂给拍死。

首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活

也可以把它当作「stand by」来用,是一个不错的桌面高颜值「配件」。

首发 我用 2 万的 iPhone Duo 指挥 8 万的 Mac Studio 干活

还有更多有意思的玩法也可以去 bilibili@光谱 主页查看。

除了炫耀,Mac Studio 能做的很多

这套价值十万元的组合,留在家里的那八万元,才是真正承担更多工作的一部分。

从苹果一开始对 Mac Studio 这个系列的定位,就能知道这不是一台面向一般消费者的电脑。

Image

创作者、开发者、工程师和各类专业人士才是这台电脑的目标使用者,强悍超凡的表现和设备端 AI 算力让 Mac Studio 几乎可以说是 AI 时代的最强桌面电脑。

笔者这次也是想要让 Mac Studio 作为一个端侧 AI 算力中心,它会负责加载模型,也会给出 Agent 需要的文件、软件和运行环境,在后台默默处理各种复杂的 Agent 任务;前台则是借助手机开展连接和控制。

Image

借助 LM Studio,笔者下载了 DeepSeek V4 Flash、Qwen3.8 27B 以及 GLM 5.3 Flash 等本地模型,再结合应用内 Local Model API 的能力,LM Studio 可以专门用来负责给出本地模型接口。

类似 LM Studio 的工具有很多,可以替代的挑选还有 Ollama、unsloth 等,unsloth 是更新比较快的本地模型部署平台,它们经常会推出不同的量化尺寸,来让更多设备都能运行这些最新的前沿开源模型。

Image

应对了本地模型接口之后,笔者这次使用 DeepSeek Harnesss 来负责规划任务、调用工具和操作文件。

DeepSeek Harness 兼容自定义 OpenAI 兼容接口,可以据此接入 LM Studio,这样模型推理和任务执行都留在了这台 Mac Studio 上。

Image

笔者使用 DeepSeek V4 Flash 跑了一个骑自行车的鹈鹕,初次生成用时 2 分 18 秒,后续笔者调整了两轮,平均输出速度在 34 token/s。

APPSO 也在准备更详细的 Mac Studio 实测,对于 AI 时代越来越贵的手机和电脑,苹果在某种程度上反而成了「性价比」的新挑选。Mac Studio 也是如此,同样 256 GB 的内存,要买一块英伟达的显卡,眼下一样售价不菲。

人们有什么建议,希望看到笔者拿 Mac Studio 做点什么整活也可以在评论区留言,笔者到时候一起「端」上来。

Image

虽然现在手机的表现也足够用来运行一小部分的本地模型,就像笔者在测试手机和电脑连接时,下载了 LM Studio 出片的 Locally,点开应用就能看到官方建议多款精选开源小模型。

Image

像是常见的 Qwen3.5、Google 的 Gemma 4,还有欧洲最强开源模型 Ministral 3,以及面壁智能推出的 MiniCPM 之类,这些模型大多是 0.8B-9B 不等,有些模型也来到了 27B,占据 4.8GB 大小。

笔者尝试了专门做端侧智能的面壁智能团队,其发布的全新旗舰多模态模型 MiniCPM-V 4.7,在手机上也能处理除文本外的视频、多图、单图等内容,总体多模态能力对标 GPT-4o。

Image

随着开源模型和闭源模型之间竞争越来越白热化,这些旗舰模型的能力也会很快就来到更小的端侧设备里。不过对 Agent 来说,模型很关键,运行环境也很关键。

就像这套组合里,Mac Studio 是中枢,它不需要移动,只需要永远在线、永远有算力。

iPhone Duo 折叠形态则让它达成了一机三用,合上是仪表盘,半折是指挥中心,展开是控制台,它不像是手机,更像是改变了笔者之前抱着出门的笔记本电脑。

尽管普通手机也能派任务、收通知,但 Duo 多出来的空间和形态,则让检查和修改更容易留在手机上完成。

Image

简单操作时,它还是一台能随手掏出来的手机;需要仔细处理时,再把显示屏展开。笔者不用为了偶尔浮现的一次复杂操作,一直带着电脑,也不必把所有事情都挤在外屏里做完。

这也是折叠形态与 Agent 工作方式契合的地方,任务可以持续执行,人按需要介入,显示屏也跟着这一次要处理的事情变化。

这套工作流,用现有的模型、Agent 和一个适配 Duo 的 App,就已经能跑起来。Siri AI 还没来,但在交互形态上,iPhone Duo 已经为 AI 时代的工作流做好了准备。

本文来源:https://www.ifanr.com/1685170
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。

The $21,499 iPhone Duo, plus the $86,749 Mac Studio, adds up to more than $100,000 for both devices.

With these two most expensive phones and computers in Apple's history, APPSO wants to build one of the most elegant "smart personal centers" of our time.

256G RAM, 1.2TB/s The memory bandwidth of the super-powerful Mac Studio stays at home and is responsible for running the model and performing Agent tasks.

Image

Just take a Duo with you when you go out, take out your phone when you need it, and you can see where the computer at home is busy, or give it some work.

For this reason, the author also made an app that is deeply adapted to the iPhone Duo. According to the three different shapes of the lid, half-fold hover, and fully unfolded, the appearance corresponds to the different factory conditions of the Agent.

Image

First of all, I arranged the local computing power of Mac Studio, and I began to work on how to make the phone connect to the computer. Just like Codex, you can assign tasks on your phone and take advantage of the local computer environment to perform a variety of complex tasks.

From the process point of view, the main problems to deal with are how the phone finds the Mac, how the task is handed over to the Mac, and how the implementation is returned to the phone.

Image

In order for the phone to connect in, another HTTPS proxy is run on the Mac to listen to the Mac's local area network address. After receiving the phone request, it is forwarded to the bridge of the machine.

If you want to connect with a cellular network outside in the future, you can let the phone and Mac join the same Tailscale private network, and then use the Tailscale Serve to give the Bridge as a private HTTPS service.

Once connected, the app checks if the service is working properly before reading the list of sessions and the contents of the selected session. The session selected by the author has a sessionId, and subsequent tasks are sent to this session.

Image

After the task enters Harness, the Agent on the Mac calls the local model, reads the workspace, executes the tool, and saves the execution record. Loss of cellular connectivity does not trigger task cancellation; tasks are still managed by Harness as long as Mac and related services continue to run.

Open the app, and there is a row of task cards on the front. I can see which Agent is running, where it is, and how long it takes. All I can do is scan it. At the top is the load summary of Mac Studio, with information such as memory usage, model name, token usage, number of running tasks, number of completed today, and so on.

Since there is such an expensive computer in the house, you must know what it is like to be busy now.

Image

The most frequent operation in this form is probably approval. If it is a sandbox environment in the cloud, you can directly change the permission of the project from "manual approval" to "approval for me" without hesitation, but if it is your own computer, it is still necessary to insure that the normal operation is automatically performed, and the key decision is asked again.

Therefore, halfway through the Agent's work, there are often situations that need to be confirmed by someone. In the closure mode, the two buttons of "confirm/reject" are made large and prominent. Just under the thumb, when and where you take out your phone from your pocket and open the app to approve.

Now, for the longer and longer tasks of the Agent, it can also be hung on the Smart Island, and the lock screen can also see where the project progress has been implemented. Almost 90% of remote interactions are closed-loop in this form.

Hands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac Studio

This is the most "only Duo can do" form of the whole machine.

Fold the Duo in half and place it on the table, which automatically cuts it up and down.

The upper half of the screen is a live broadcast, and the progress of Agent execution will automatically scroll the full trajectory like a video monitor, covering tool calls, model inference, and so on.

Hands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac Studio

The bottom half of the screen is the operation, the trackpad, the terminal input box, the approval button, and a voice dispatch button. It is placed on the table in a half-folded state, which is the most natural gesture for voice input.

Measuring the most pleasant situation of this form, like chasing the drama, Duo stood in the corner of the table and watched the machine at home render the film one by one.

The fully expanded Duo is close to a small tablet, and all the "too compact and lazy to pull out the computer on the phone" operations all fall into this form.

The layout is like a classic double column on a computer, with a task queue on the left and a full session of tasks selected on the right. Agent-generated products, including rendered previews, organized tables, edited videos, and web pages, have plenty of room for previews, and can be annotated more conveniently if you are dissatisfied, so that it can be redone.

Hands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac Studio

What's more interesting is the split-screen operation of opening multiple apps. Just like the author swipes the photo album to see a reference picture, you can directly drag in the task card, which is the input of new tasks.

Secondly, two Agent sessions can indicate side by side. The left side looks at the generation progress of the webpage, and the right side changes the prompt of the other task.

Instead, the three shapes become the same as the three apps. Cover the conclusion, hover over the screen, and expand to the details.

The author saw the approval request in the closed state, expanded it to see the details, and then folded it back and stuffed it back into the pocket after approval. The information density can follow the display area, which is the most essential difference between the folding screen client and the ordinary mobile phone app.

Image

Based on several iPhone Duo shapes, I also vibe coding a few other interesting apps, so that folding this shape to play its real value.

Just like the best posture of this fish, outsiders think that they are reading an in-depth report, in fact, the folded large screen part is surfing the Internet.

The author can even customize the cover and set it to a different textbook cover, such as advanced mathematics.

There is also a game corresponding to the shape and appearance of the iPhone Duo. When the display is closed, the full screen cockroaches are slapped to death.

It can also be used as a "stand by" and is a good desktop high-value "accessory".

Image

There are many more fun ways to play on the bilibili @ spectrum homepage.

This set of combinations worth 100,000 yuan, the 80,000 yuan left at home, is really part of taking on more work.

From the beginning of Apple's positioning of the Mac Studio series, it can be seen that this is not a consumer-oriented computer.

Creators, developers, engineers, and professionals of all kinds are the target users of this computer, and the powerful performance and AI computing power on the device make Mac Studio almost the most powerful desktop computer in the AI era.

The author also wants to use Mac Studio as an AI computing power center on the end. It will be responsible for loading the model, and will also give the files, software and operating environment required by the Agent, and silently handle various complex Agent tasks in the background. The front desk is connected and controlled by the mobile phone.

Hands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac Studio

With the help of LM Studio, I downloaded local models such as DeepSeek V4 Flash, Qwen3.8 27b, and GLM 5.3 Flash. Combined with the ability of the Local Model API in the application, LM Studio can be specifically used to give the local model interface.

There are many tools similar to LM Studio, and alternative choices include Ollama, unsloth, etc. Unsloth is a fast updating local model deployment platform. They often introduce different quantitative sizes to allow more devices to run these latest cutting-edge open source models.

After dealing with the local model interface, I used DeepSeek Harnesss to plan tasks, call tools, and manipulate files.

DeepSeek Harness is compatible with custom OpenAI-compatible interfaces and can access LM Studio accordingly, so that model inference and task execution are left on this Mac Studio.

The author used DeepSeek V4 Flash to run a bicycle pelican, which took 2 minutes and 18 seconds to generate for the first time. Subsequently, the author adjusted two rounds, and the average output speed was 34 token/s.

Image

APPsO is also preparing more detailed Mac Studio tests, and Apple has become a new choice for "cost performance" in the era of increasingly expensive mobile phones and computers in the AI era. Mac Studio is the same, the same 256 GB of memory, to buy a Nvidia graphics card, the same price is expensive.

People have some suggestions. You can also leave a comment in the comments area if you want to see the author do something with Mac Studio. The author will "end" it at that time.

Although the performance of the mobile phone is now enough to run a small number of local models, just like when I tested the connection between the mobile phone and the computer, I downloaded the Locally produced by LM Studio and clicked on the app to see the official suggestions for a variety of small open source models.

Such as the common Qwen3.5, Google's Gemma 4, Ministral 3, the strongest open source model in Europe, and MiniCPM launched by Face Wall Intelligence. Most of these models range from 0.8B-9B, and some models have also come to 27b, occupying 4.8GB in size.

The author tried a face wall intelligence team specializing in end-side intelligence. The new flagship multi-modal model MiniCPM-V 4.7 released by the team can also handle video, multi-graphs, single-graphs and other content in addition to text on mobile phones, and the overall multi-modal capability is benchmarked against GPT-4o.

As the competition between open-source and closed-source models intensifies, the capabilities of these flagship models will soon come to smaller end-side devices. However, for the Agent, the model is critical, and the operating environment is also critical.

Just like in this set, Mac Studio is the hub, it doesn't need to move, it just needs to be always online and always has computing power.

Image

the folding shape of the iPhone Duo makes it a machine with three uses. It closes the dashboard, half folds the command center, and expands the console. It is not like a mobile phone, but more like a laptop that the author held before going out.

Although ordinary phones can also send tasks and receive notifications, the extra space and shape of Duo make it easier for inspections and modifications to be left on the phone.

When it is simple to operate, it is still a phone that can be taken out at will; when it needs to be handled carefully, the display is expanded. The author does not have to carry a computer all the time for an occasional complex operation, and does not have to squeeze everything into the outer screen to finish.

This is also where the folding form fits in with the Agent's way of working. Tasks can continue to be performed, people intervene as needed, and the display also changes with the things to be dealt with this time.

This workflow is already running with existing models, Agents, and an app for Duo. Siri AI is not here yet, but in interactive form, the iPhone Duo is ready for the workflow of the AI era.

This article is from: https://www.ifanr.com/1685170 For copyright issues, please write to: 13951431913#139.com or submit a review on the right.

ImageImageHands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac StudioHands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac StudioHands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac StudioHands On: Using a 20,000-Yuan iPhone Duo to Drive an 80,000-Yuan Mac StudioImageImageImageImageImageImageImageImage

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