Before RTX Spark laptops ship, NVIDIA takes PAIR out of beta and lays out a Windows-on-Arm developer stack

NVIDIA used October 13, three days before the first RTX Spark laptops go on sale, to publish a long list of software that it says is ready for the platform. RTX Spark is NVIDIA's Arm-based Windows PC chip; NVIDIA describes it as offering up to 128GB of unified memory and 1 petaflop of AI compute. Laptop availability on Friday, October 16 was announced last week by NVIDIA and Microsoft.

The post mixes items shipping today with items scheduled for later this month. For developers who run models locally, the parts that matter are a local inference router, the state of the Python and CUDA toolchain on Windows on Arm, and a new open model.

PAIR leaves beta

NVIDIA PAIR, short for Personal AI Router, is a free, open-source tool that discovers computers on a local network and sends independent inference requests to whichever machine has the right engine, model and spare capacity. It launched in beta in early September with support for Ollama and LM Studio, according to NVIDIA's IFA post.

NVIDIA says the official release adds llama.cpp support, direct model downloads from Hugging Face, configurable inference engine settings and RTX Spark support. It runs on Windows, Linux and macOS and exposes Ollama-compatible and OpenAI-compatible endpoints, so existing apps and agents can point at it without code changes.

The PAIR 1.0.0 playbook is explicit about its limits, and they are worth knowing before you plan around it: PAIR does not combine GPU memory, does not merge GPUs into one larger device, and does not split a model or a single request across machines. It routes whole requests. If a model does not fit on one of your computers, PAIR will not make it fit. It also accepts requests only from the machine where it is installed.

The Windows-on-Arm toolchain

Arm-based Windows machines have long had gaps in the Python and GPU tooling that ML developers rely on. NVIDIA says that, in partnership with Microsoft and the community, the following are now available:

• The first PyTorch-CUDA build for Windows on Arm.
• The first Python build that is native by default on Windows on Arm, with Python 3.15.
• A native Conda installer for Windows on Arm with more than 2,000 packages.
• According to NVIDIA, 97% of the top 15,000 Python wheels are ready; the rest are still being worked on.

That last figure is the one to check against your own dependency list. A package missing an Arm64 wheel, or a compiled extension that has not been rebuilt, can still block a project even when the headline libraries work.

Agents, models and llama.cpp

NVIDIA also says OpenClaw and GitHub Copilot now support Microsoft Execution Containers (MXC), an operating-system-level sandbox for agents, and that Hermes Agent will add it by the end of October. It released MCP servers for several of its own apps, including Broadcast and PAIR. On the llama.cpp side, a Windows-native companion app called Llama Windows launches today, and llama.cpp now supports what NVIDIA calls decision models. NVIDIA's claim that these models let agents choose actions "up to 200x faster than traditional agents" is a vendor figure with no published methodology in the post.

The model to watch is Nemotron 3.5 Bolt 74B-A9B, due October 15: a mixture-of-experts model with 74 billion total and 9 billion active parameters, packed into 24GB with 2-bit quantization. NVIDIA says it will be fully open, including training data, weights and recipes, and runnable at launch through llama.cpp, vLLM, Ollama, Unsloth and LM Studio. Quality claims for it come only from NVIDIA so far.

What is not in the post

NVIDIA did not publish pricing for RTX Spark laptops in this post, and its performance figures (up to 2x faster Adobe workflows, the 200x decision-model figure) are its own. Several items, such as the Hermes integrations and Blender ray reconstruction, are scheduled for later dates rather than shipping today. Independent testing will follow once laptops reach buyers on October 16.