NVIDIA RTX Spark ships today with up to 128GB unified memory and 1 petaflop of AI compute, as OpenShell and MXC lock down local agents

Start with the positioning and specifications NVIDIA published. The company's blog calls RTX Spark a new beginning for the PC, a reinvention combining the capabilities of a gaming PC, a creative studio and a personal AI supercomputer into a single superchip, with up to 128GB of unified memory and 1 petaflop of AI compute, delivering the full stack of RTX and CUDA technologies to creators, developers and gamers while unlocking a new personal-computing paradigm through local AI agents. On timing, NVIDIA says it and Microsoft announced last week that RTX Spark laptops would be available on Friday, October 16, and that in the run-up NVIDIA and its ecosystem partners are rolling out new software features and performance optimisations to expand what local AI can do on the machine. Note that the blog gives no pricing and no table breaking down memory and compute across configurations.

The second point is the agent line, which carries the densest information in the whole list. NVIDIA says RTX Spark greets users at first boot with a Getting Started screen to choose and install their agent app, and that these agents come with new secure sandboxing technology from NVIDIA OpenShell and MXC, Microsoft Execution Containers. The division of labour is explicit: MXC delivers operating system-level security, while OpenShell makes it simpler for agents to adopt MXC and adds advanced capabilities such as credential protection and security monitoring. NVIDIA states that Hermes Agent will integrate granular OpenShell controls and MXC by the end of the month, and that OpenClaw and GitHub Copilot now support MXC. On routing, the blog describes NVIDIA Switchyard as an open source model router library that lets agents intelligently route queries between local and cloud LLMs, so users can minimise cloud usage while still obtaining near-cloud results; Hermes will add Switchyard later in October, while Perplexity Computer and GitHub Copilot have integrated their own routing solutions. After installation, apps including GitHub Copilot, Hermes Agent, OpenClaw, Perplexity, Pi and WorkBuddy recommend an optimised local model and settings for RTX Spark and set it all up in a few clicks.

The third point is the concrete list of models and creative apps. On models: Nemotron 3.5 Bolt 74B releases on Thursday, October 15, which NVIDIA describes as purpose-built for agentic AI on RTX Spark; NVIDIA PAIR moves out of beta and adds llama.cpp support; and Nemotron 3.5 Bolt E4B 4GB, the third-generation small language model for NVIDIA ACE AI game characters, arrives on Friday, October 23. On creative and display: Adobe's top apps are getting up to 2x performance, with NVIDIA and Adobe collaborating to rearchitect Photoshop for GPU-accelerated compositing and Premiere for playback, rendering and effects; NVIDIA Broadcast 2.3 brings video super resolution, video frame generation, offline file processing and agent Model Context Protocol controls; RTX Video frame generation and HDR are available now in ComfyUI; and DLSS Ray Reconstruction arrives in Blender 5.3 in November. On development and gaming: PyTorch, Python, CUDA and all CUDA libraries are ready for RTX Spark, llama.cpp adds decision-model support and ships a Windows app; a 2x faster DLSS 5 AI model arrives on October 16 with a new Performance Mode for DLSS Frame Generation available today via the NVIDIA App beta; and the latest RTX Remix update with DLSS 5 support releases on October 23.

The boundaries should be listed one by one. First, every performance and speedup figure above comes from NVIDIA's own launch messaging; these are vendor claims, not independently reproduced third-party results, and statements such as 2x or 1 petaflop especially need their test conditions spelled out. Second, part of the list is roadmap rather than shipping capability: Hermes Agent's OpenShell and MXC integration is billed for the end of the month, Switchyard for later in October, DLSS Ray Reconstruction in Blender 5.3 for November, and Adobe's 2x performance as coming soon. Third, the blog publishes no price for RTX Spark and no breakdown of configurations, and does not say whether the Getting Started experience is preinstalled or downloaded. Fourth, two trend figures deserve separate reading: NVIDIA says local LLM downloads are already 5x higher than all of last year and describes local LLMs as having crossed the intelligence threshold to power agents, citing Qwen 3.8-Flash-Next catching up to frontier cloud models released just a few months ago. These are NVIDIA's own framings without third-party statistics attached.