CONFIRMED: This is based on NVIDIA’s own blog post of 13 October 2026. All performance claims are NVIDIA’s and have not been independently tested. Nemotron 3.5 Bolt is announced for Thursday 15 October and is not available yet.
NVIDIA has set out the software it wants running on its RTX Spark laptops, which go on sale on Friday 16 October. The focus is AI agents that run on the computer itself rather than in the cloud [1].
The headline is a new open AI model, Nemotron 3.5 Bolt 74B. NVIDIA also says its PAIR tool, which shares AI work across your home computers, is leaving beta today [1].
RTX Spark is NVIDIA’s new laptop chip, announced last week with Microsoft. NVIDIA says it offers “up to 128GB of unified memory and 1 petaflop of AI compute” [1].
The new model: Nemotron 3.5 Bolt 74B (ANNOUNCED, out 15 October)
NVIDIA says: “NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB, available Thursday, Oct. 15, is a new LLM purpose-built for running local agents on RTX Spark” [1].
It describes it as “a 74B parameter, 9B-active-parameter mixture-of-experts (MoE) model, packed into 24GB using 2-bit quantization” [1]. In plain English:
- Mixture of experts (MoE). The model has 74 billion parameters (the numbers it learned in training), split into many specialist parts. For each word it writes, only about 9 billion are used. That makes it much faster than a model that uses everything every time.
- 2-bit quantization. Normally each number in a model takes 16 bits or more of memory. Squeezing them to 2 bits shrinks the model so it fits in 24GB. That usually costs quality, so NVIDIA says it trained the model knowing it would be squeezed (“quantization-aware training”) to keep quality up [1].
NVIDIA claims it is “faster than most popular dense models while smarter than most popular MoE models with a similar memory footprint” [1]. That is NVIDIA’s claim. The model is not out yet, and TSN has seen no independent tests.
NVIDIA says the model is “fully open source — including training data, weights and recipes”, so developers can adapt it. It will run at launch through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and directly in agents such as Hermes and OpenClaw [1].
PAIR and Switchyard: routers for your AI
A model router decides where a piece of AI work should run.
- PAIR, which NVIDIA calls a “personal AI router”, is “officially moving out of beta today”. It is free and open source. It finds compatible PCs on your home network, installs models and spreads AI work across them. This release adds llama.cpp support and direct model downloads from Hugging Face. It works on Windows, Linux and macOS [1].
- Switchyard is “an open source model router library”. It lets an agent send easy questions to a free local model and harder ones to a paid cloud model, so you use the cloud less. NVIDIA says Hermes will add it later in October [1].
Easier setup and safer agents
NVIDIA says setting up a local model is now a few clicks in GitHub Copilot, Hermes Agent, OpenClaw, Perplexity, Pi and WorkBuddy. The apps suggest a model and settings suited to RTX Spark [1].
For safety, agents can run in a sandbox, a fenced-off area that limits what they can touch. This uses NVIDIA OpenShell together with Microsoft Execution Containers (MXC). NVIDIA says “OpenClaw and GitHub Copilot now support MXC”, and Hermes Agent will add it by the end of the month [1].
NVIDIA also claims “local LLM downloads are already 5x higher than all of last year” [1]. It did not publish the data behind this figure.
Also announced
- Adobe: “Up to 2x performance is coming soon” to Adobe’s top apps; optimisations for AI features in Photoshop and Premiere are in beta now (NVIDIA’s claim) [1].
- NVIDIA Broadcast 2.3: out now, with an MCP server that lets local AI agents change effects and settings [1].
- ComfyUI: two new RTX Video nodes, for frame generation and HDR [1].
- Developers: “The first PyTorch-CUDA build for Windows on Arm” [1].
- Games: a small model, Nemotron 3.5 Bolt E4B 4GB, for AI game characters, releases on 23 October [1].
What this means
- The model is not out yet. Nemotron 3.5 Bolt 74B is due on 15 October. Until then, its speed and quality are only NVIDIA’s claims.
- It is meant to be fully open, including training data, which is rarer than open weights alone.
- PAIR is available now, free and open source, and not only for RTX Spark: it also runs on Linux and macOS.
- The aim is less cloud. NVIDIA wants agents to run on your own laptop, and only call on the cloud when needed.
Sources
- NVIDIA Blog (Gerardo Delgado), “All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark”, 13 October 2026. https://blogs.nvidia.com/blog/local-ai-rtx-spark-laptops-software-updates/
