NVIDIA and Microsoft announced on October 13 that Windows laptops built on RTX Spark will go on sale on Friday, October 16. The chip made its debut at Computex in June, when only hardware specs and a vague “this fall” window were offered. This time the companies have added a firm date and a long list of software.
The hardware figures are unchanged: up to 128GB of unified memory and 1 petaflop of AI compute, with the CPU and GPU built into a single “superchip” running Windows on Arm. In its blog post, NVIDIA calls it “a new beginning for the PC.” The post does not say which brands will ship first or what the machines will cost; that will have to wait until retailers list them on the 16th.
Developer tooling comes first
The most common complaint about Windows on Arm has been that software lags behind. This time NVIDIA spelled out the developer layer in detail:
- PyTorch, Python, CUDA and all CUDA libraries are available, including the first PyTorch-CUDA build for Windows on Arm;
- Python 3.15 runs natively on Windows on Arm, and the native Conda installer ships with more than 2,000 packages;
- 97% of the 15,000 most-downloaded Python wheels are already ready;
- llama.cpp becomes a backend for Windows ML, and a Llama Windows app is being released.
The blog does not say which wheels make up the remaining 3%, or whether any of them are widely used libraries.
Agent apps and local routing
On the application side, GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy get a simplified local install flow and two layers of sandboxing: NVIDIA’s OpenShell and Microsoft’s execution container, MXC. OpenClaw and GitHub Copilot already support MXC, while Hermes will not add OpenShell and MXC until the end of October.
NVIDIA is also open-sourcing a model-routing library called Switchyard, which sends each request to either a local or a cloud model. PAIR, a personal AI routing tool, is leaving beta; together with llama.cpp support it will work on Windows, Linux and macOS. GitHub Copilot uses a mechanism called HydraFusion for local model routing, and Microsoft’s MAI 1.1 Coding Flash can run locally.
For creative software, Adobe says Photoshop, Premiere and Substance 3D Painter are optimized for RTX Spark with up to 2x performance; the AI-feature optimizations in Photoshop and Premiere are still in beta. Blender 5.3 beta already supports the chip, and ComfyUI has added RTX video frame interpolation and HDR nodes. On the gaming side, the 2x DLSS 5 model arrives with NBA 2K27 on October 16 and reaches RTX Remix on October 23.
Doing the memory math
The flagship model for this launch is Nemotron 3.5 Bolt 74B-A9B, released on October 15: 74 billion total parameters, 9 billion active per pass, 2-bit quantization, roughly 24GB on disk. The weights, training data and training recipe are all open, and it runs on llama.cpp, vLLM, Ollama, Unsloth and LM Studio. NVIDIA describes it as aiming for the best balance between speed and intelligence.
A rough calculation: 74 billion parameters at 2 bits come to about 18.5GB of weights alone, and with quantization metadata and some layers kept at higher precision, NVIDIA’s 24GB figure checks out. In 128GB of unified memory the model itself takes up less than a fifth, leaving room for a long-context KV cache or a second small model acting as a router. By the same estimate, a 4-bit build would need about 37GB or more, which would still fit. Judging by size alone, the 24GB tier looks aimed at lower-spec machines with less memory, though NVIDIA has not explained why it chose 2-bit.
The blog gives no benchmark comparison showing how much capability 2-bit quantization costs against higher-precision versions, and there are no third-party evaluations yet. That is the first thing someone will need to test once the laptops ship.
Another model, Nemotron 3.5 Bolt E4B, is only about 4GB. It is the third generation of small models for NVIDIA ACE game characters and is due on October 23. NVIDIA also said downloads of local large models are now five times last year’s full-year total; the blog does not explain how that figure is counted.
Price is the biggest unknown for these machines. At the June launch NVIDIA said only that the first wave would be positioned at the high end, with versions carrying less memory to follow. Once listings go up on the 16th, the gap between the 128GB flagship and the lower-memory configurations will show how much extra it costs to run a 70B-class model locally on a laptop.
Sources: NVIDIA official blog (on-sale date, software ecosystem and Nemotron 3.5 Bolt specs); CocoLoop; Computex 2026 RTX Spark launch materials (memory and compute figures).