THE SHORT VERSION
- ASUS announced a 64GB Ascent GX10 configuration alongside the existing 128GB version; the pressroom page is dated October 5 while the release dateline reads October 2.华硕为现有 128GB 版本新增 64GB Ascent GX10 配置；新闻页面日期为 10 月 5 日，正文新闻稿日期则写 10 月 2 日。
- ASUS says 64GB unified memory is sized for efficient 30–35B open models; its 1-petaflop headline is tied to theoretical FP4 performance using sparsity.华硕称 64GB 统一内存适合高效的 300 亿至 350 亿参数级开放模型；1 petaflop 是采用稀疏计算的理论 FP4 性能宣传。
- The device runs NVIDIA DGX OS on a 20-core Arm CPU and GB10 GPU, has a fixed 1TB SSD configuration and limited internal upgrades, and is built for AI development rather than Windows desktop use.该设备使用 NVIDIA DGX OS、20 核 Arm CPU 和 GB10 GPU，配备华硕地区规格列出的固定 1TB SSD 且内部升级有限，定位为 AI 开发机而非 Windows 台式电脑。
The 64GB model lowers memory, not the platform64GB 版本改变的是内存容量，不是平台
ASUS added a 64GB configuration of the Ascent GX10 alongside its existing 128GB model. The ASUS Pressroom page is dated October 5, 2026, while the release dateline inside it says October 2; the announcement describes both versions as compact systems built around the NVIDIA DGX Spark platform. The new configuration is aimed at developers, researchers, and teams starting local inference, not at a conventional mini PC buyer looking for a Windows office computer.华硕在现有 128GB 版本之外增加了 64GB Ascent GX10。华硕 Pressroom 网页日期为 2026 年 10 月 5 日，新闻稿正文日期则写着 10 月 2 日；公告称两款都是基于 NVIDIA DGX Spark 平台的紧凑型系统。新配置面向开始本地推理的开发者、研究人员和团队，不是想买一台 Windows 办公电脑的普通迷你主机用户。
The main difference is unified memory capacity. ASUS says the 64GB version is sized for efficient open models in the 30–35-billion-parameter range, while the 128GB version targets larger local models and heavier workloads. This is a manufacturer use-case claim, not a guarantee that every model in that size range will fit comfortably. Model weights, quantization, context length, runtime, and the operating system all consume memory, so the workload matters more than a parameter-count label alone.主要区别是统一内存容量。华硕称，64GB 版本适合高效的 300 亿至 350 亿参数级开放模型；128GB 版本则面向更大的本地模型和更重的工作负载。这是厂商对适用场景的描述，并不保证每个同级模型都能顺畅装入。模型权重格式、量化、上下文长度、运行环境和操作系统都会占用内存，因此工作负载比参数量标签更重要。
Parameter count is a starting point, not a capacity promise参数量只是起点，不是容量保证
A 35-billion-parameter model does not require only 35GB of memory. The exact footprint depends on how weights are represented, how long a prompt context is kept, and how much memory is reserved for the system and active tools. A quantized model can take less space than a full-precision one, but quality, speed, and compatibility may change. ASUS's release mentions models such as Qwen3.8-27B as a fit example; the model, software version, and task should be checked together before a purchase.一个 300 亿至 350 亿参数模型不一定只需要相同数值的 GB 内存。实际占用取决于权重表示方式、保留的提示词上下文长度，以及系统和当前工具预留的内存。量化模型占用更少，但质量、速度与兼容性可能改变。华硕新闻稿以 Qwen3.8-27B 为适用示例；购买前应把具体模型、软件版本和任务一起核对。
For a personal prototype that runs one moderate model at a time, 64GB can be a lower entry point into the GB10 platform. A lab that needs several models loaded, larger context windows, or more concurrent agents may value the 128GB version. Neither option changes the need to test actual software: a model can load and still respond too slowly for a workflow. Buyers should list the model files, context sizes, framework, and expected users before translating the headline into a memory requirement.如果个人原型工作只需一次运行一个中型模型，64GB 可以降低进入 GB10 平台的门槛。实验室若要同时装载多个模型、使用更大的上下文或支持更多并发代理，可能更看重 128GB。两种配置都不能免去软件测试：模型能够启动，不代表它能按工作需要及时返回结果。买家应先列出模型文件、上下文大小、运行框架和预期使用人数，再把宣传转成内存需求。
The performance headline depends on a special calculation峰值性能数字依赖特定计算方式
ASUS advertises up to one petaflop of AI performance from the GB10 Grace Blackwell superchip. Its footnote ties that number to theoretical FP4 performance using sparsity. FP4 is a low-precision numeric format used in some AI workloads; sparsity can skip certain zero-valued operations. A petaflop figure under those conditions is not a measured output rate for every model, nor a prediction of how many words a local agent returns each second.华硕宣传 GB10 Grace Blackwell 超级芯片最高可达 1 petaflop AI 性能。脚注将数值限定为采用稀疏计算的理论 FP4 性能。FP4 是部分 AI 工作负载使用的低精度数值格式；稀疏计算可以跳过某些零值运算。以这些条件得出的 petaflop 数字，不等于每种模型的实测速度，也不能直接预测本地代理每秒能返回多少文字。
The independent TechRadar hands-on, published January 27, examined the original 128GB GX10 rather than the new 64GB configuration. Its author describes a 20-core Arm CPU and Blackwell GPU in a 150mm-square enclosure and notes that the machine runs NVIDIA DGX OS based on Ubuntu. Those details help explain the product's purpose, but the review does not independently test the 64GB option. The practical way to compare it with a cloud service or desktop GPU is to run the same model, quantization, context length, and workload, then compare speed and total system cost.独立媒体 TechRadar 于 2026 年 1 月 27 日发布的上手评测实际接触的是原版 128GB GX10，而不是新 64GB 配置。其作者描述了 150 毫米方形机身中的 20 核 Arm CPU 和 Blackwell GPU，并指出这台机器运行基于 Ubuntu 的 NVIDIA DGX OS。这些细节有助于理解产品定位，但该评测没有独立测试新 64GB 版本。要比较它与云服务或桌面 GPU，应使用相同模型、量化方式、上下文长度和工作负载，再比较速度和系统总成本。
Two 64GB units can be clustered, with extra setup两台 64GB 设备可以组网，但需要额外配置
ASUS says the GX10 64GB includes NVIDIA ConnectX-7 networking and can be linked with a second unit through the NVIDIA Sync Cluster Assistant. The company advertises up to 128GB of pooled memory, up to two petaflops of AI compute, and up to 1.7 times the throughput of a single 128GB unit. These are vendor figures for a defined clustered setup. They do not mean every application automatically combines two boxes or doubles its speed; software must support distributed execution and the network has to be configured.华硕称 GX10 64GB 内置 NVIDIA ConnectX-7 网络接口，可通过 NVIDIA Sync Cluster Assistant 与第二台设备连接。公司宣传两台可提供最高 128GB 池化内存、最高 2 petaflops AI 算力，并达到单台 128GB 设备约 1.7 倍吞吐量。这些数字对应厂商定义的集群场景。它们并不意味着任何应用都能自动合并两台设备，或速度必然翻倍；软件需支持分布式运行，网络也要正确配置。
The two-box route can make sense if a developer wants to start with one machine and later add capacity. It also means buying another desktop, power supply, cable, and desk space, then managing a multi-node software environment. For an occasional local model test, a single 64GB unit may be easier to operate. For an organization that already has AI tools and staff, pooling memory could be useful—but the use case should justify the second unit rather than treating scalability as free.如果开发者想先从一台开始、以后再增加容量，两台设备的方案可能有用。这也意味着要再买一台主机、电源、线缆并预留桌面空间，同时管理多节点软件环境。偶尔测试本地模型的用户可能更容易维护单台 64GB 设备；已经配备 AI 工具和人员的组织，才更容易发挥池化内存的价值。应让工作需求支撑第二台设备的成本，而不是把可扩展能力当作免费的升级。
Treat it like a compact workstation, not an upgradeable tower应把它看成紧凑工作站，而非可升级塔式电脑
ASUS lists a 150-by-150-by-51mm enclosure weighing 1.48 kilograms, with a 240W power supply, 10GbE LAN, Wi-Fi 7, HDMI 2.1, and four USB-C ports. There is no USB-A port, so a mouse, keyboard, external drive, or security key may require a hub or adapter. The chassis is small, but it is not a low-power stick PC; the external supply and sustained AI workload should be considered when choosing a shelf, UPS, or desk location.华硕给出的机身尺寸为 150×150×51 毫米，重量 1.48 千克，配有 240W 电源、10GbE 网口、Wi-Fi 7、HDMI 2.1 和四个 USB-C 接口。没有 USB-A 接口，因此鼠标、键盘、外置硬盘或安全密钥可能需要扩展坞或转接头。机箱虽小，但它并非低功耗电视棒；选择书架、UPS 或桌面位置时，还要考虑外置电源和持续运行的 AI 工作负载。
The ASUS Canada specifications page lists a 1TB M.2 PCIe 4.0 SSD for the published 64GB configuration. Its family table notes a single M.2 slot and up to 4TB, but ASUS says users cannot change the installed SSD and that opening the chassis may void the warranty. The 64GB column itself lists 1TB, so the family-level maximum is not a user-upgrade promise. TechRadar's 128GB review also calls out limited internal expansion; its observation is not a teardown of the 64GB SKU.华硕的加拿大地区规格页面为公布的 64GB 配置列出 1TB M.2 PCIe 4.0 SSD。该系列规格表还注明只有一个 M.2 插槽、最高支持 4TB；但华硕同时表示用户不能更换已安装的 SSD，拆开机箱可能导致保修失效。64GB 一栏本身列出的仍是 1TB，因此系列级最高容量不等于用户可自行升级。TechRadar 对 128GB 版本的评测也提到内部扩展有限，但那不是对 64GB SKU 的拆机检查。
The operating system is built around AI development系统软件围绕 AI 开发设计
The GX10 ships with NVIDIA DGX OS and a developer-focused software stack that ASUS lists as PyTorch, Jupyter, Ollama, OpenShell, and NemoClaw. Its software and hardware fit people who run local AI frameworks, build agent workflows, or experiment with models on an Arm/Linux environment. It is not a plug-in replacement for a Windows gaming or productivity desktop. Before ordering, check that required Python packages, container images, drivers, and IDE extensions support the machine's architecture and software stack.GX10 随附 NVIDIA DGX OS 和开发者软件栈，华硕列出 PyTorch、Jupyter、Ollama、OpenShell 和 NemoClaw。它适合运行本地 AI 框架、构建代理工作流，或在 Arm/Linux 环境中试验模型。它不是可直接替代 Windows 游戏或办公主机的电脑。下单前应确认所需 Python 程序包、容器映像、驱动和 IDE 扩展支持这台设备的芯片架构与软件栈。
Local inference can keep prompts and model execution on a system the owner controls, which may help with confidential development data. That benefit applies to workflows actually running locally. If an agent calls a cloud API, uses a hosted search service, or synchronizes files remotely, data may still leave the machine. A buyer should map those network calls and software defaults instead of treating “local AI” as a blanket privacy guarantee. ASUS's product positioning supports local development; the user's application and network configuration determine what stays local.本地推理能让模型运行和提示词留在用户控制的系统上，在处理保密开发数据时可能有帮助。但这只适用于真正本地运行的工作流。如果代理调用云端 API、托管搜索服务或远程同步文件，数据仍可能离开设备。买家应梳理软件会发起哪些网络连接，而不是把“本地 AI”视为全面隐私保证。华硕产品定位支持本地开发；具体留存方式仍取决于应用与网络配置。
Who should consider the 64GB option谁适合考虑 64GB 版本
The 64GB GX10 is for a developer or research group that can name its open models, frameworks, and local workloads. ASUS says to contact a local representative for availability and pricing, so the October announcement does not provide a universal 64GB price. TechRadar's earlier 128GB review reported retailer prices for that larger configuration; those figures do not establish the price of the new SKU. Ask for the exact memory, SSD, warranty, and local delivery terms before comparing it with the NVIDIA DGX Spark or a GPU workstation.64GB GX10 适合能够明确列出开放模型、运行框架和本地工作负载的开发者或研究团队。华硕称购买者应联系当地代表查询供应与价格，因此这次 10 月公告没有提供统一的 64GB 全球售价。TechRadar 较早评测的 128GB 型号曾列出零售商价格，但不能用来推断新 SKU 的价格。询价时应确认内存、SSD、保修和本地交付条件，再与 NVIDIA DGX Spark 或 GPU 工作站比较。
For a general home user who needs web browsing, games, office apps, or video editing, the Linux-based software stack and limited internal expansion make the GX10 an unusual fit. For an AI-curious builder who wants local inference and can troubleshoot Linux, it offers a compact path into the DGX Spark platform. The decision is less about whether one petaflop sounds impressive and more about whether 64GB is enough for the exact model and context, whether the software supports Arm, and whether the workload justifies the system's specialist price.对需要网页浏览、游戏、办公软件或视频剪辑的一般家庭用户来说，Linux 软件栈和有限的内部扩展意味着 GX10 用途特殊。如果 AI 开发者希望本地推理并能处理 Linux 环境，它则提供了进入 DGX Spark 平台的紧凑途径。决定因素不在于“1 petaflop”听上去有多大，而在于 64GB 是否适合所选模型与上下文、软件是否支持 Arm，以及这项工作负载是否值得购买一台专业计算设备。
Sources & reporting limits来源与核查范围
Follow Computing & Hardware to find related stories on your next visit.关注电脑与硬件，下次访问时阅读相关文章。
