NVIDIA positions RTX Spark 128GB for local agents

NVIDIA on May 31 used GTC Taipei at Computex to introduce RTX Spark as a new Windows PC class built around local AI agents, not just gaming or content creation. The company said the system combines up to 128GB of unified memory with up to 1 petaflop of FP4 AI performance, a spec sheet it framed as sufficient for “personal agents” that run on-device rather than in the cloud. (blogs.nvidia.com) NVIDIA’s product page says RTX Spark will ship in slim laptops and small desktop PCs from partners including ASUS, Dell, HP, Lenovo, Microsoft and MSI. The announcement matters because NVIDIA is now describing the PC as a host for software that works continuously in the background. (blogs.nvidia.com) Its RTX Spark page says the machines are “built to run personal AI agents 24/7,” while the company blog says the hardware is meant to support agents that can write code, generate assets and handle multi-step tasks locally. ### Why is NVIDIA emphasizing 128GB of unified memory? NVIDIA put 128GB of unified memory at the center of the pitch. The company said that memory pool, shared across CPU and GPU resources, is meant to handle the processing demands of on-device agents and let developers prototype, fine-tune and run current models locally. (blogs.nvidia.com) The DGX Spark hardware guide shows what NVIDIA means by that configuration in practice. The desktop system uses 128GB of LPDDR5X unified system memory, a 20-core Arm CPU, 6,144 CUDA cores and 273 GB/s of memory bandwidth, according to NVIDIA’s documentation. The same guide says DGX Spark supports models up to 200 billion parameters, or 405 billion in a dual-Spark configuration. (nvidia.com) ### What does NVIDIA mean by “local agents”? NVIDIA’s May 31 blog described local agents as software that can interact with applications, automate repetitive processes and manage multistep tasks while running on the device. The company said security and privacy are part of that argument, because local execution reduces the need to send workloads out to cloud services. (blogs.nvidia.com) Hermes Agent is one of the examples NVIDIA is using to make that case. In a May 13 post, NVIDIA said Hermes is designed for “always-on local use,” can access local files and applications, and uses contained sub-agents to break larger tasks into smaller ones. NVIDIA also said Hermes saves learned skills over time and is built as an orchestration layer for persistent, on-device agents rather than one-off task execution. (docs.nvidia.com) ### Where do Hermes and OpenClaw fit into this? NVIDIA said Hermes Agent and OpenClaw will integrate OpenShell and Microsoft’s security primitives into new Windows applications. The May 31 post also said the NemoClaw blueprint is being expanded across GeForce RTX, RTX PRO, RTX Spark, DGX Spark and DGX Station, with streamlined installers and support for Hermes Agent. (blogs.nvidia.com) The company has been building that message for several weeks. NVIDIA’s Hermes post said the framework is provider- and model-agnostic, optimized for local use and suited to smaller context windows and concurrent workloads on local hardware. (blogs.nvidia.com) ### Is this only about a new chip, or about a software stack too? CUDA is part of the hardware pitch, but NVIDIA is pairing the chip with software and developer tooling. The RTX Spark product page says CUDA runs natively on the platform, while the May 31 blog says OpenShell is coming to Windows and that llama.cpp and vLLM are getting inference and multi-GPU optimizations for agentic workloads. (blogs.nvidia.com) This fall is the next concrete milestone NVIDIA has given. NVIDIA said the broader set of local-agent updates, including OpenShell on Windows and partner software changes, will arrive alongside RTX Spark systems later in 2026. (blogs.nvidia.com) (nvidia.com) (blogs.nvidia.com)