NVIDIA published on October 13, 2026, an extensive overview of the software that will accompany the launch of RTX Spark, the new laptop platform the company presents as a reinvention of the PC: a single superchip combining the capabilities of a gaming PC, a creative studio and a “personal AI supercomputer.” Last week, NVIDIA and Microsoft confirmed that the first RTX Spark laptops will be available on Friday, October 16. According to the company, the platform offers up to 128GB of unified memory and 1 petaflop of AI compute.
The common thread of the announcement is AI agents running on the device itself. NVIDIA says agent popularity skyrocketed this year, with projects such as OpenClaw and Hermes topping GitHub charts, and that local models have crossed the intelligence threshold needed to power them. The company says local language model downloads are already five times higher than all of last year, and cites Qwen 3.8-Flash-Next as an example of a local model catching up with cloud models from just a few months ago.
The centerpiece is Nemotron 3.5 Bolt 74B, a model built specifically to run local agents on RTX Spark, available Thursday, October 15. It is a mixture-of-experts model with 74 billion parameters, 9 billion of them active at each step, packed into 24GB using 2-bit quantization with quantization-aware training techniques to retain quality. NVIDIA says it is faster than most popular dense models and smarter than most popular mixture-of-experts models with a similar memory footprint.
Like the rest of the Nemotron family, Bolt 74B will be fully open: NVIDIA will publish the training data, weights and recipes so developers can fine-tune it and build their own agents. It will run from day one on llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and will be directly accessible from agents such as Hermes and OpenClaw. The company also announced Nemotron 3.5 Bolt E4B, a small 4GB model for video game characters with NVIDIA ACE, arriving October 23.
The second major item is PAIR, an open-source “personal AI router” that officially leaves beta. The tool automatically discovers and connects compatible machines on a local network, installs models and inference engines and distributes workloads among them, so users can tap the combined power of several PCs. The official release adds llama.cpp support, direct downloads from Hugging Face, customizable inference engine settings and RTX Spark support. PAIR works on Windows, Linux and macOS.
Agent security features prominently. Agents installed on RTX Spark will be sandboxed using NVIDIA OpenShell and Microsoft Execution Containers (MXC). MXC provides operating-system-level security, while OpenShell makes it easier for agents to adopt MXC and adds credential protection and security monitoring. According to NVIDIA, OpenClaw and GitHub Copilot already support MXC, and Hermes Agent will integrate granular OpenShell controls and MXC by the end of the month. It is a direct response to concerns about agents accessing sensitive files and services.
Setup is also simplified. From first boot, machines will show a welcome screen to choose and install an agent app. GitHub Copilot, Hermes Agent, OpenClaw, Perplexity, Pi and WorkBuddy will recommend an optimized local model and suitable settings and configure everything in a few clicks. For those who want to combine local and cloud, NVIDIA offers Switchyard, an open library that routes queries between local and cloud models to reduce reliance on remote services; Hermes will add it later in October.
The development ecosystem is another front. Because this is a Windows platform on Arm architecture, NVIDIA and Microsoft have worked to make common tools work at launch: the first PyTorch build with CUDA for Windows on Arm, the first Python release native by default on that architecture — Python 3.15 — and the first native Conda installer with more than 2,000 packages. NVIDIA says 97% of the 15,000 most-used Python packages are already ready.
There is also news for decision models, a recent category of models that return probabilities over predefined options rather than free text. Llama.cpp and Unsloth add support for them on RTX Spark and GeForce RTX GPUs, and llama.cpp debuts Llama Windows, a native app for downloading and running local models without the command line. Microsoft, for its part, will offer coding models such as MAI 1.1 Coding Flash locally and add llama.cpp as a Windows ML backend.
The announcement also includes creative and gaming improvements. Adobe promises up to twice the performance in Photoshop, Premiere and Substance 3D, with optimizations already available in beta; NVIDIA Broadcast 2.3 adds video super resolution and frame generation, processing of recorded files and an MCP server so agents can control the app; and DLSS 5 with neural rendering arrives on launch day, with a twice-as-fast model debuting in NBA 2K27. Blender will add DLSS Ray Reconstruction in November.
The bet has market implications. For years, advanced AI has been concentrated in large data centers, with pay-per-use costs and dependence on connectivity. A laptop able to run 74-billion-parameter models with sandboxed agents offers an alternative for users and companies that prioritize privacy, latency or predictable costs. At the same time, it puts NVIDIA in more direct competition with Apple, Qualcomm, AMD and Intel in the AI PC arena.
Open questions remain that only independent testing will settle: the real-world performance of a 2-bit quantized model compared with higher-precision equivalents, laptop battery life under heavy AI workloads, final device prices and the maturity of Windows on Arm compatibility for less common software. It also remains to be seen whether sandboxing with MXC and OpenShell holds up in practice against the bypass attempts observed with other agents.
With launch set for October 16, NVIDIA is trying to ensure RTX Spark arrives with a full ecosystem, not just hardware. Optimized open models, tools to spread workloads across machines, agents sandboxed by default and support for the main development tools paint a clear vision: that the PC’s next big leap is running capable AI assistants directly on the user’s desktop.
