Nvidia RTX Spark Windows laptops go on sale with 128GB unified memory and 1 petaflop of FP4

RTX Spark pairs a Blackwell RTX GPU with a Grace CPU in one superchip, with up to 128GB of unified memory shared between graphics and local models. Laptops come from ASUS, Dell, HP, Lenovo, Microsoft and MSI. Nvidia ships its Nemotron 3.5 Bolt 74B-A9B model, tuned for tool calling and multistep tasks, for supported configurations, plus guided installers for agents including GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy that need no command line. Agent permissions are handled by Nvidia's OpenShell runtime and Microsoft Execution Containers. A feature called PAIR lets one RTX Spark machine serve inference to Windows, Linux and macOS computers on the same home network. The launch follows this week's release of Nemotron 3.5 Super VL and the 2-bit Bolt model aimed at local agents. For AI builders, this sets a concrete hardware target for local-first agents on Windows, where large models and agent runtimes can share one memory pool. Teams selling privacy-sensitive or offline agent products can test against it, and the secure-runtime pieces hint at where permission models for desktop agents are heading.