刚刚,首批搭载英伟达 RTX Spark 超级芯片的 Windows 笔记本电脑开售了。
APPSO 也提前拿到华硕 ProArt 创16 RTX Spark,这款笔记本 16 寸的 64GB 统一内存版本售价为 28999 人民币,128GB 统一内存版本则是 37999 人民币。


这是英伟达第一次以这样的方式进入 Windows PC 的核心硬件平台。以前的 Windows 电脑,大多数是使用 Intel 或者 AMD 的处理器,英伟达只是向电脑厂商给出显卡。

现在,英伟达开始把自己的 ARM CPU、Blackwell RTX GPU 和统一内存结合起来,推出一套可以运行 Windows 的完整计算平台。

在感受了一段时间后 APPSO 给人们挖出了更多细节,作为 AI PC 跟 Mac 比较如何,是不是可以当「理财机型」来等着它增值……
把大模型装进一台最薄 12.9 mm 的电脑
笔者拿到的是 64GB 统一内存版本,由于是统一内存的形式,实际可用的 GPU 显存大约在 46.7 GB 左右,LPDDR5X 的内存类型,让它的内存带宽速度峰值理论在 300GB/s,但这也能让它运行像 Qwen 3.8 270 亿参数这样的模型。

另一维度,过去的 Windows ARM 架构的笔记本电脑,主要想回答的是「能不能像 MacBook Air 一样轻薄、省电、电池续航长」。
而 RTX Spark 的 ARM64 架构则让这台电脑能取得类似 Apple Silicon Mac 的高能效和统一内存长处,又保留了 Windows 的应用生态和 NVIDIA 的 CUDA 能力。

在电池续航上,这台华硕 ProArt 创16 RTX Spark 的电池是 99.9Wh,笔者找了一个本地 270 亿参数模型持续生成文本和推理,拔掉电源,能看到电池的放电功率大约在 86.5W 左右,而剩余的电量几乎是在 2 分钟的时间里,就从 98.5Wh 掉到了 96.1Wh。


假设整机平均功耗是 80W,99.9Wh 理论电池续航也能达到约 1 小时 15 分钟。也就是说,如果你正在持续运行一个 27B 或更大的模型,可以在不插电的情况下坚持约一个小时。

对于一台能在本地运行大型 AI 模型的笔记本,一台最薄 12.9mm,重约 1.77kg 的笔记本,这意味着即使离开电源,它还能继续承担一段时间的高负载计算任务。

电池续航之外,这台华硕 ProArt 创16 RTX Spark 笔记本电脑配备了多元的外部接口,3 个 USB4 Type-C,每个最高 40Gbps,兼容视频输出和 USB PD 供电。
1 个 USB-A 接口,使用 USB 3.2 Gen 2,最高 10Gbps;1 个 HDMI 2.1 FRL 接口用来连接外部表明器,以及 1 个 SD Express 7.0,用于高速存储卡和 1 个 3.5mm 音频接口。


这些接口最大的意义,也是让它可以成为一台移动工作站:外接高速硬盘、4K 表明器、相机存储卡,以及各种创作设备。
64GB 统一内存,能让 AI 干多少活
前几天,在微软的特别活动上,他们把 Windows PC 分成了开箱即用的迷你台式电脑、具备混合智能,本地和云端 AI 的 Copilot + PC,以及能利用云端和本地 AI 开展创作的 Builder PC。
这台华硕 ProArt 创16 RTX Spark笔记本电脑正是位于 Builder PC,即创作者个人电脑的类别下,而眼下最热门的创作,自然是用 AI 来开展,生成游戏、3D 建模,还有简单的网页应用等,AI 都必须参与其中。

对华硕 ProArt 创16 RTX Spark来说,引入 AI 的方式之一就是本地部署,因为这样才能充分利用英伟达 RTX Spark 的能力。
笔者下载了 LM Studio,并安装了 Qwen 3.8 27B 的模型。将思考深度调整为极高,问一句「你好」,这台笔记本总共花了 18 秒,能看到模型读取、处理输入内容的速度 Prompt Speed 是 615.7 token/s,生成 token 速度在 14.3 token/s,首 token 的时间是 2.83s。

而当笔者尝试要求它做一些网页项目时,例如为一个 AI PC 的活动外观网页。

可以看到生成 Token 的速度还是稳定在 14 token/s 左右,首 token 的时间是 12.05s,总的用时为 609.67s。

对于 64GB 的统一内存来说,30B 左右的模型大概就是这类电脑的舒适区。
Qwen3 32B(NVFP4)的推理速度在每秒 55–60 tokens 上下,日常对话、文案写作、代码辅助完全流畅;喂入长文档做总结时,十万 token 级的上下文也没有显著衰减。

WorkBuddy,我的智能体现在 0 元/月
除了单独的本地部署,这台 Windows 电脑最有意思的地方是和 WorkBuddy 的配合。
打开我 MacBook Air 上的 WorkBuddy,能看到在设置的模型规格里,只有自定义模型这一栏。
而在这台华硕 ProArt 创16 RTX Spark 上,多出来了一个 Beta 能力,笔者可以直接下载并管理本地运行的 AI 模型。
点击下载之后,电脑会自动开始下载模型,并应对好本地部署对应的规则,笔者只需要启动,并让它常驻后台,回到 WorkBuddy 的任务界面,就可以挑选该本地模型开展对话。
之前的 Three.js 屋顶游戏,现在借助 WorkBuddy 也能达成,虽然这个成效看着比不上当下最前沿的模型,但用来做一些初步的项目验证,观察可行性还是可以试试。
整个任务用时是 13m59s,比较使用 WorkBuddy 的云端模型,时间可能还更少。
也能看到任务管理器里,表现部分,GPU 的使用率开始上升到 100%,风扇转动的声音同样显著。
虽然在 WorkBuddy 内没有直接告诉笔者这款本地模型是来自哪家模型厂商,不过和它简单的聊天,就能从模型的文件头部里找到相关信息。
这款模型是腾讯优化过的本地模型,500 亿参数的 MoE 模型,模型大小为 16GB,使用 CUDA 13 加速,在 GPU 上的推理速度能做到更加流畅。
针对腾讯优化过的本地模型的能力,笔者也能找到更合适这款模型的任务,配合 WorkBuddy 这套 Harness 的能力,一些日常的简单办公,像是数据分析、数据可视化、报告总结等繁琐型工作,都会有不错的表现。
将来随着软硬件的配合,以及当前开源模型超越闭源,且都在做更快更小的 Flash 模型趋势下,在 WorkBuddy 内可直接下载的本地模型也肯定不会只有这一个。
当前笔者也可以借助 LM Studio 本地模型 API 的方式,将 LM Studio 内下载的本地模型,以 API 的形式加载到 WorkBuddy 内。
加载成功之后,在 WorkBuddy 内就能表明对应的自定义模型。
使用同样的方式,也可以将本地模型加载到其他你常用的 Harness 工具里,像是 DeepSeek Harness。
算力很强,Windows 也在变得更好
如果前面 WorkBuddy 的感受让笔者觉得,在 Windows 上使用本地 AI 已经足够简单,但针对 Windows 开发,又总是会有各种各样的难题。
这也是 Windows、Linux 和 macOS 三大平台使用者经常纠结的原因。一维度很多开源开发工具最初就是围绕 Unix/Linux 工作流构建的,例如 Bash 脚本、Git、OpenSSL、编译器和各种命令行工具,在 Linux/macOS 环境中往往使用相似的调用方式。
另一维度,Linux 同样也会遇到驱动、平台库和不同发行版之间的兼容难题;macOS 也可能缺少 Python 包的 Apple Silicon build 版本,还可能需要安装 Xcode Command Line Tools。
好在现在笔者看到的是,Windows 一直在应对这些难题,从前几天的 Windows 更新,正在推进的 Microsoft Execution Containers MXC 项目,能为执行文件操作、调用工具的 AI 智能体给出更安全、可控的运行环境。
在本地模型维度,微软也宣布为 Windows ML 引入 llama.cpp 兼容,让开发者可以借助 Windows 的 AI 运行时调用更多开源模型。
以及之前就有的 WSL 给出 Linux 开发环境,到 Windows Terminal、WinGet,持续完善的 ARM64 原生开发工具,Windows 已经比过去更合适开发者使用。
同时,Windows 日常使用感受也在更新,新版 Windows Search,除了搜索文件和应用,还可以让平台执行对应操作,不必再进入设置菜单逐项查找。
Copilot 也准备加入更多本地能力,在取得使用者许可后,它可以读取电脑里的相关文件、执行平台操作,或者调用本地模型完成任务。
所以,AI 能成为下一次换机的理由吗
以前,一台配备大容量 NVIDIA GPU 的 AI 工作站,往往意味着体积不小的主机、独立显卡和持续供电的桌面环境。
而现在,英伟达已经把一套可以运行大型本地模型、兼容 CUDA 的计算平台,放进了一台不到 1.8kg 的 Windows 笔记本里。
笔者可以把它带去办公室,放在咖啡店的桌子上,甚至拔掉电源,让一个 27B 模型继续运行。
但这台电脑也很清楚地暴露了现阶段本地 AI 的另一面:硬件已经可以运行大型模型,软件却还没有完全做好让每个人都从容使用它的准备。
WorkBuddy 这样的机型正在减少部署门槛,更多智能体也开始尝试直接利用本地算力。另一维度,软件适配、模型能力、表现与电池续航之间的取舍,仍旧会影响最终感受。
所以,如果你每天都需要运行本地模型、开发 AI 应用,或者处理不方便上传云端的资料,这一代华硕 ProArt 创16 RTX Spark 笔记本电脑确实给出了一个值得考虑的新挑选。
但如果你平时只是使用 ChatGPT、Gemini 等云端 AI 工具,现在还没有必要为了本地运行一个大模型,就专门换掉手里的笔记本。
眼下对于个人电脑来说,总体的趋势一定是往越大内存、越强算力、越多 AI 的方向发展。
当 AI 智能体越来越多地参与笔者每天的工作,电脑本身的计算能力就会重新变得关键。以前笔者可能只需要一台电脑连接云端 AI,以后也可以让一部分任务直接留在电脑里完成。
本文来源:https://www.ifanr.com/1684682
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
The first Windows laptops with Nvidia RTX Spark superchips have just gone on sale.

APPSO also received the Asus ProArt 16 RTX Spark in advance. The 16-inch 64GB unified memory version of this notebook is priced at RMB 28,999, and the 128GB unified memory version is RMB 37,999.
This is the first time that NVIDIA has entered the core hardware platform of a Windows PC in this way. Previous Windows computers, most of which used Intel or AMD processors, Nvidia only gave computer manufacturers graphics cards.

Now, Nvidia is combining its own arm CPU, Blackwell RTX GPU, and unified memory to deliver a complete computing platform that runs Windows.
After feeling it for a while, APPSO dug out more details for people. As an AI PC, how is it compared to a Mac? Can it be a "financial model" to wait for it to add value...

The author got the 64GB unified memory version. Because it is in the form of unified memory, the actual available GPU graphics memory is about 46.7 GB. The memory type of LPDDR5X makes its memory bandwidth speed peak theory at 300GB/s, but this also allows it to run models like Qwen 382.7 billion parameters.

In another dimension, in the past, Windows arm-based laptops mainly wanted to answer the question, "Can you be as thin and light as a MacBook Air, save power, and have a long battery life?"
RTX Spark's ARM64 architecture allows the computer to achieve energy efficiency and unified memory advantages similar to Apple's Silicon Mac, while preserving the Windows application ecosystem and NVIDIA's CUDA capabilities.

In terms of battery life, the battery of this Asus ProArt 16 RTX Spark is 99.9Wh. I found a local 27 billion parameter model to continuously generate text and inference, unplugged the power supply, and I could see that the discharge power of the battery was about 86.5W, and the remaining power fell from 98.5Wh to 96.1Wh in almost 2 minutes.
Assuming the average power consumption of the whole machine is 80W, the theoretical battery life of 99.9Wh can also reach about 1 hour and 15 minutes. That said, if you are running a 27b or larger model continuously, you can stick around for about an hour without plugging in.

For a notebook that can run large AI models locally, a notebook that is the thinnest at 12.9mm and weighs about 1.77 kg means that it can continue to take on high load computing tasks for a while, even without power.
In addition to its battery life, this Asus ProArt Tron 16 RTX Spark laptop is equipped with multiple external interfaces, 3 USB4 Type-C, each up to 40Gbps, compatible with video output and USB PD power supply.

1 USB-A connector for USB 3.2 Gen 2 up to 10Gbps; 1 HDMI 2.1 FRL connector for external indicator and 1 SD Express 7.0 for high speed memory card and 1 3.5mm audio connector.

The biggest significance of these interfaces is that they can become a mobile workstation: external high-speed hard drives, 4K indicators, camera memory cards, and various authoring devices.
The other day, at a special event at Microsoft, they divided Windows PCs into mini-desktops that work right out of the box, Copilot + PCs with hybrid intelligence, on-premises and cloud AI, and Builder PCs that leverage both cloud and on-premises AI for creation.

This Asus ProArt Creative 16 RTX Spark laptop is located in the category of Builder PCs, that is, creators' personal computers, and the most popular creations right now are naturally carried out with AI, generating games, 3D modeling, and simple web applications. AI must be involved.
For the Asus ProArt Gen 16 RTX Spark, one of the ways to introduce AI is to deploy it locally, as this leverages the power of the Nvidia RTX Spark.

I downloaded LM Studio and installed a model of Qwen 3.8 27b. Adjust the thinking depth to extremely high, ask "Hello", this notebook took a total of 18 seconds, you can see that the speed of the model reading and processing the input content Prompt Speed is 615.7 token/s, the token generation speed is 14.3 token/s, and the time of the first token is 2.83 s.
And when I try to ask it to do some web project, for example, for the active appearance of an AI PC.

You can see that the speed of generating tokens is still stable at about 14 tokens/s, the time of the first token is 12.05 s, and the total time is 609.67s.

For 64GB of unified memory, a model around 30b is probably the comfort zone for such computers.
The inference speed of Qwen3 32b (NVFP4) is around 55–60 tokens per second, and daily conversations, copywriting, and code-assisted are completely smooth. When feeding long documents for summarization, the context of 100,000 tokens is not significantly attenuated.

Apart from individual on-premises deployments, the most interesting thing about this Windows computer is its work with WorkBuddy.
Open WorkBuddy on my MacBook Air and you can see that in the model specification settings, only the custom model column is displayed.

On this Asus ProArt Gen 16 RTX Spark, there is an additional Beta capability, and I can directly download and manage AI models running locally.
After clicking Download, the computer will automatically start downloading the model and deal with the rules corresponding to the local deployment. The author only needs to start it and let it reside in the background. Return to the Task interface of WorkBuddy, and you can select the local model to start the conversation.

The previous Three.js rooftop game can now be achieved with the help of WorkBuddy. Although this performance is not comparable to the current cutting-edge model, it can be used to do some preliminary project verification and observe the feasibility.

The entire task takes 13m59s, which may be less time than using WorkBuddy's cloud model.
You can also see that in the task manager, the performance part, the utilization rate of the GPU begins to rise to 100%, and the sound of the fan turning is also significant.

Although I was not directly told in WorkBuddy which model manufacturer this local model is from, you can find relevant information from the model's file header by simply chatting with it.
This model is a Tencent-optimized local model, a 50 billion-parameter MoE model, the model size is 16GB, using CUDA 13 acceleration, the inference speed on the GPU can be more smooth.

For the ability of the local model optimized by Tencent, I can also find more suitable tasks for this model. With the Harness ability of WorkBuddy, some simple daily work, such as data analysis, data visualization, report summary and other tedious tasks, will perform well.
In the future, with the combination of software and hardware, and the current open source model surpassing closed source, and both doing faster and smaller Flash models, the local model that can be downloaded directly within WorkBuddy will certainly not be the only one.

Currently, I can also use the LM Studio Native Model API to load the local model downloaded in LM Studio into WorkBuddy as an API.

Once loaded, the corresponding custom model can be indicated within WorkBuddy.
In the same way, you can also load the local model into other Harness tools you commonly use, such as DeepSeek Harness.

If the previous experience of WorkBuddy makes me feel that using native AI on Windows is simple enough, but there are always various difficulties in developing for Windows.
This is why users of the three major platforms, Windows, Linux and macOS, often struggle. One-dimensional Many open source development tools were originally built around Unix/Linux workflows, such as Bash scripts, Git, OpenSSL, compilers, and various command-line tools, often using similar invocation methods in Linux/macOS environments.

In another dimension, Linux will also encounter compatibility problems between drivers, platform libraries, and different distributions; macOS may also lack the Apple Silicon build version of the Python package, and may require the installation of Xcode Command Line Tools.
Fortunately, what I see now is that Windows has been coping with these difficulties. From the Windows Update of the past few days, the Microsoft Execution Containers MXC project is being promoted, which can provide a more secure and controllable operating environment for AI agents that perform file operations and call tools.

In the local model dimension, Microsoft also announced the introduction of llama.cpp compatibility for Windows ML, allowing developers to call more open source models with the help of Windows' AI runtime.

And the previous WSL gives the Linux development environment, to Windows Terminal, WinGet, the continuous improvement of the ARM64 native development tools, Windows has been more suitable for developers than in the past.
At the same time, the daily experience of Windows is also being updated. The new version of Windows Search, in addition to searching for files and apps, can also allow the platform to perform corresponding operations without having to enter the settings menu to find each item.

Copilot is also ready to add more local capabilities, after obtaining the user's permission, it can read the relevant files in the computer, perform platform operations, or call the local model to complete the task.
Previously, an AI workstation with a high-capacity NVIDIA GPU often meant a large host, a discrete graphics card, and a continuously powered desktop environment.

Now, Nvidia has put a CUDA-compatible computing platform that can run large local models into a Windows laptop that weighs less than 1.8kg.
I can take it to the office, put it on the coffee shop table, or even unplug it and let a 27b model continue to run.

But the computer also clearly exposes the other side of local AI at this stage: the hardware can already run large models, but the software is not yet fully ready for everyone to use it.

Models like WorkBuddy are reducing deployment thresholds, and more agents are starting to try to leverage local computing power directly. Another dimension, the trade-off between software adaptation, model capability, performance and battery life, still affects the final feeling.
So if you need to run on-premises models, develop AI applications, or handle data that's inconvenient to upload to the cloud every day, this generation of Asus ProArt 16 RTX Spark laptops does offer a new option to consider.

But if you usually only use cloud-based AI tools such as ChatGPT and Gemini, it is not necessary to run a large model locally, so replace the notebook in your hand.
At present, for personal computers, the overall trend must be towards larger memory, stronger computing power, and more AI.

As AI agents are more and more involved in the author's daily work, the computing power of the computer itself will become critical again. In the past, I may only need a computer to connect to the cloud AI, and in the future, I can leave some tasks directly on the computer.
This article is from: https://www.ifanr.com/1684682 For copyright issues, please write to: 13951431913#139.com or submit a review on the right.