NVIDIA has published a roundup of software updates timed to the launch of RTX Spark laptops, which NVIDIA and Microsoft say go on sale Friday, Oct. 16. The hardware pairs up to 128GB of unified memory with 1 petaflop of AI compute. Most of the post is about gaming and creative apps, but several items matter to teams building agents.
A model for local agents
Nemotron 3.5 Bolt 74B arrives Thursday, Oct. 15. It is a mixture-of-experts model with 74B parameters, of which 9B are active at a time. Two-bit quantization packs it into 24GB. NVIDIA positions it as faster than popular dense models and smarter than popular MoE models of similar memory size; those comparisons are its own and the source includes no benchmark data. The training data, weights and recipes are all open source, so teams can fine-tune it. At launch it runs through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and agents such as Hermes and OpenClaw can use it directly.
Security and routing
Agents installed on RTX Spark come with sandboxing from NVIDIA OpenShell and Microsoft Execution Containers (MXC). MXC provides operating-system-level security. OpenShell makes it easier for agents to adopt MXC and adds credential protection and security monitoring. OpenClaw and GitHub Copilot already support MXC, and Hermes Agent is due to add granular OpenShell controls and MXC by the end of the month.
NVIDIA Switchyard, an open-source model router library, lets an agent send queries to either a local or a cloud model. The aim is to use the cloud less while getting similar results. Hermes will add it later in October. Perplexity Computer and GitHub Copilot have built their own routing.
PAIR and decision models
NVIDIA PAIR, a free open-source tool that finds PCs on a local network and spreads AI workloads across them, leaves beta. It installs models and inference engines, now supports llama.cpp and Hugging Face downloads, and runs on Windows, Linux and macOS.
llama.cpp and Unsloth now support decision models, which NVIDIA says let an agent choose its next action, for example by classifying a request, up to 200x faster than traditional agents. llama.cpp also gets a Windows app that runs from the system tray, so models can be downloaded and run without the command line.
What is not in the source
The post gives no laptop prices and no independent tests. Its claim that local model downloads are 5x higher than all of last year is NVIDIA's own. The gaming, Adobe, Blender and ComfyUI items are outside what this note covers.