At GTC Berlin NVIDIA added three product updates to its local AI hardware portfolio: RTX Spark brings fully local personal agents to Windows laptops, DGX Spark ships a 64 GB configuration with built‑in clustering for up to eight nodes, and DGX Station will gain a Windows variant later this year. The announcements also include software extensions such as Underdog Desktop Agent, NVIDIA Sync updates, and early‑access Cloud Functions.
Hardware Updates for Local AI
RTX Spark is positioned as a high‑end PC platform with up to 128 GB unified memory, 1 petaflop of AI compute, and support for personal agents running entirely on device. The platform is available from major OEMs (ASUS, Dell, HP, Lenovo, Microsoft, MSI) with an unlocked bootloader on many models to enable custom firmware or OS changes.
DGX Spark 64 GB systems start shipping on 23 Oct. They can host models such as Nemotron 3.5 Bolt 74B, Qwen 3.8 27B, and Gemma 4 26B‑A4B. Built‑in ConnectX‑7 networking and NVIDIA Sync Cluster Assistant allow manual configuration of clusters up to eight units, delivering roughly 1 TB of unified memory across the cluster.
DGX Station, aimed at larger teams, already offers 748 GB coherent memory and up to 20 petaflops. A Windows‑based version is slated for December, extending the same hardware capabilities to Windows‑native workflows and professional applications.
Software and Tooling Changes
Underdog Desktop Agent, built for RTX Spark, runs the Husky inference engine locally, ensuring that user data never leaves the PC. The agent is available for immediate download.
NVIDIA Sync received two updates: support for up to eight DGX Spark nodes and a Model Launcher UI that automates model and container retrieval, then connects the model to OpenCode for troubleshooting and documentation.
The October DGX Spark software update adds a Wi‑Fi Assistant that can switch networks while the appliance runs, and a set of DGX Spark Skills that embed workflow guidance directly into OpenCode.
New open‑source reference workflows for robotics are published for DGX Station, and NVIDIA Cloud Functions is announced for early access on the same platform. Recursive Flint, a local coding model, is now available on all NVIDIA local hardware.
Operational and Security Implications
Running models locally reduces latency and eliminates outbound data transfers, which is relevant for privacy‑sensitive workloads. However, the unlocked bootloader on many RTX Spark laptops introduces a potential attack surface; organizations should assess firmware signing policies before enabling custom images.
Clustered DGX Spark deployments rely on ConnectX‑7 networking and NVIDIA Sync; practitioners must plan for network segmentation, monitoring, and configuration management to avoid unintended exposure between nodes.
Integration with enterprise management tools (Absolute Security, Splunk, Microsoft Intune, Perforce, Puppet, Chef, Tanium) adds a layer of centralized control for DGX Station, but also requires alignment with existing security policies and credential management processes.
Model Launcher automates container pulls, which may introduce supply‑chain considerations. Verifying container signatures and maintaining a trusted registry become part of the operational checklist.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Engineers should evaluate whether the local compute capacity of RTX Spark or DGX Spark meets the latency and data‑privacy requirements of their agentic AI workloads. For teams needing multi‑user access, the eight‑node DGX Spark cluster offers a path to shared model serving without cloud dependency.
Security teams must incorporate bootloader policies, network segmentation, and supply‑chain verification into the deployment workflow. Operations teams should prototype the Model Launcher and Wi‑Fi Assistant to understand their impact on device maintenance and incident response.
Finally, keep an eye on the upcoming Windows DGX Station release and the early‑access NVIDIA Cloud Functions, as they will broaden the range of Windows‑centric applications that can leverage on‑premise AI models.
