NVIDIA staggers RTX Spark agent and creative updates from Oct. 15 into November

NVIDIA's agent model, Nemotron 3.5 Bolt 74B, is built for agentic AI on RTX Spark ^[6]. At one byte per weight, 74 billion parameters take 74GB. At half a byte, they take 37GB ^[16]. On the 128GB top configuration that leaves 54GB or 91GB for the operating system, the agent app and the context cache ^[17]. The "up to" in the spec means smaller configurations exist ^[2], and they leave less.

NVIDIA calls the petaflop figure AI compute ^[2]. The post does not state the numeric precision behind it, the memory bandwidth, or the weight format of the 74B model. Single-stream token generation tends to be limited by how fast weights stream out of memory, so a compute peak says little about tokens per second. For the number to transfer to an agent loop, the loop would need to be compute-bound, as long-prompt processing is, and the model would need to run at the precision the peak assumes.

MXC delivers operating-system-level security, and OpenShell makes it simpler for agents to adopt MXC while adding credential protection and security monitoring ^[11]. OpenClaw and GitHub Copilot already support MXC ^[4]. In our view that is the right layer for an agent that can run commands and hold credentials. Whether it holds against hostile input is a question for outside testing. The only account so far is NVIDIA's own post.

The defaults matter more than the feature list. At first boot a "Getting Started" screen has the user choose and install an agent app ^[10]. The app then recommends a local model and settings for RTX Spark and configures them in a few clicks ^[9]. NVIDIA says GitHub Copilot is already optimized for RTX Spark ^[12]. Most installs will run whatever the recommendation picks.

NVIDIA says users want local models for free iteration and cloud models for frontier intelligence ^[13]. It introduces Switchyard, which it describes as open source, in that same passage ^[15]. Whatever the default split between local and cloud turns out to be, it becomes a cost and data-handling setting on every install.

NVIDIA heads its list "New and upcoming RTX Spark updates" ^[14]. The 74B model arrives one day before the laptops ^[18]. The Oct. 23 releases arrive seven days after them ^[19]. Blender gets DLSS Ray Reconstruction in November ^[7]. The Adobe "up to 2x" is a ceiling on the paths NVIDIA rearchitected, Photoshop compositing and Premiere playback and rendering. It transfers to a project only if that stage dominates your time ^[20].