Both versions are aligned block by block, in reading order: title, the essentials, then paragraph by paragraph. Where the translation merged or split a paragraph, the matching cell stays empty — we never pair two passages by guesswork.
拔掉电源，这台16英寸的华硕ProArt 创16仍在跑一个270亿参数的大模型。爱范儿的测试里，电量从100%降到85%左右只用了十分钟。这台机器重约1.77kg，最薄处12.9mm，是英伟达RTX Spark平台的首批产品之一。10月16日，首批搭载它的机型正式开售。
Unplugged, this 16-inch ASUS ProArt Creator 16 is still running a 27-billion-parameter large model. In Ifanr's testing, the battery dropped from 100% to about 85% in just ten minutes. The machine weighs about 1.77 kg and is 12.9 mm thick at its thinnest point, making it one of the first products built on Nvidia's RTX Spark platform. The first models went on sale October 16.
过去，本地跑大模型的瓶颈通常不在显卡算力，而在显存。爱范儿指出，哪怕买到顶配的5080显卡，显存也就16GB，参数稍大的模型权重连加载都困难。RTX Spark的思路不同：CPU和显卡共用同一块内存，也就是所谓的统一内存，显卡可以直接调用整块内存池，最高达到128GB。芯片由英伟达与联发科合作设计CPU部分，并沿用英伟达的CUDA生态，即开发者常用的GPU编程工具链，原本在台式机上跑的代码可以直接搬到笔记本上。
In the past, the bottleneck for running large models locally was usually not GPU compute but video memory. Ifanr points out that even with a top-end RTX 5080, the video memory is only 16GB, which makes it hard to even load the weights of larger models. RTX Spark takes a different approach: the CPU and GPU share the same memory, known as unified memory, so the GPU can directly access the entire memory pool, up to 128GB. The CPU portion of the chip was designed by Nvidia in collaboration with MediaTek, and it uses Nvidia's CUDA ecosystem, the GPU programming toolchain commonly used by developers, so code that ran on desktops can be moved directly to laptops.
实测数字因机型和模型而异。爱范儿用Ollama加Open WebUI运行Qwen3.5 35B-A3B（Q4量化）时，生成速度达到59.65 Token每秒，Token是模型处理文字的基本单位。换成参数更密集的Qwen3.8 27B，预热后为20.98 Token每秒。在华硕机型上用LM Studio运行Qwen 3.8 27B，生成速度约为14.3 token/s。这些是媒体的现场测试结果，并非独立机构的统一评测。
Measured results vary by model and machine. When Ifanr ran Qwen3.5 35B-A3B (Q4 quantization) with Ollama and Open WebUI, generation speed reached 59.65 tokens per second. Tokens are the basic units in which models process text. Switching to the denser Qwen3.8 27B, the speed was 20.98 tokens per second after warm-up. On the ASUS machine, running Qwen 3.8 27B with LM Studio, generation speed was about 14.3 tokens per second. These are the media's own hands-on tests, not unified evaluations by independent institutions.
对需要在客户现场部署模型、随身调试硬件、或处理不便上传云端的资料的人，这套平台已经有了直接用途。智能体软件WorkBuddy可以下载本地模型，离线状态下部分任务也能跑完；英伟达还表示，其NVIDIA PAIR功能可把这台电脑的AI推理能力提供给同一家庭网络中的其他电脑。不过，WorkBuddy目前还不支持端云自动切换，用户需要自己选择用本地模型还是云端模型。
For people who need to deploy models at customer sites, debug hardware on the go, or work with materials that are inconvenient to upload to the cloud, the platform already has practical uses. The WorkBuddy agent software can download local models, and some tasks can be completed offline. Nvidia also says its NVIDIA PAIR feature can share the computer's AI inference capability with other computers on the same home network. However, WorkBuddy does not yet support automatic switching between on-device and cloud models, so users must choose between local and cloud models themselves.
价格和生态是最大的保留意见。Surface Laptop Ultra起步价21988元，且入门款只配24GB内存和512GB存储。游戏方面，微软和英伟达正在与开发商逐个打通反作弊系统，微软称《使命召唤》将于2027年登陆RTX Spark平台。对只用ChatGPT、Gemini等云端工具的普通用户，爱范儿的结论很直接：暂时没有必要为了本地运行大模型换掉手里的笔记本。
Price and ecosystem are the biggest reservations. The Surface Laptop Ultra starts at 21,988 yuan, and the entry model comes with only 24GB of memory and 512GB of storage. On gaming, Microsoft and Nvidia are working with developers one by one to get anti-cheat systems running, and Microsoft says Call of Duty will arrive on the RTX Spark platform in 2027. For ordinary users who only use cloud tools such as ChatGPT or Gemini, Ifanr's conclusion is direct: there is no need for now to replace their current laptop just to run large models locally.