RTX Spark laptops for local LLMs: which memory tier to buy, and what the first week of testing found

NVIDIA’s RTX Spark laptops went on sale Friday, October 16, and compact desktops follow in November. NVIDIA is pitching them as local-AI machines: an Arm-based Windows chip with a Blackwell GPU and up to 128GB of unified memory. If you’re deciding whether one belongs on your desk instead of a Mac or a GPU tower, the useful questions are which memory tier to get, what software runs on it today, and how local agents actually behave.

Key facts:

• Chip. RTX Spark pairs a Blackwell RTX GPU with up to 6,144 CUDA cores and an up-to-20-core Grace CPU connected at 600 GB/s. NVIDIA rates it at 1 petaflop of FP4 AI compute with up to 128GB of unified memory. (Source: NVIDIA)
• Prices. The 20-core / 6,144-CUDA-core N1X laptops The Verge tested run from $3,199.99 (32GB HP OmniBook Ultra) to $6,999.99 (Asus ProArt P16). Configs with the 18-core / 5,120-CUDA-core chip start at $2,599.99 with 24GB. (Source: The Verge)
• Software at launch. NVIDIA says PyTorch, Python, CUDA and all CUDA libraries are available. Its Nemotron 3.5 Bolt 74B model runs through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, and llama.cpp has released a Windows app. (Source: NVIDIA)

Memory tier is the real decision

The Verge flagged one limit you should know before you buy. On a system with 32GB or less, Hermes Agent says there isn’t enough memory for its recommended local Qwen models. It then either routes you to the cloud or leaves you to find a smaller model yourself.

Two more details matter at the low end. Microsoft says it holds back some of the unified memory so Windows stays responsive. NVIDIA’s Windows app also has a slider that sets how much memory the GPU gets, which The Verge says is most useful on 24GB and 32GB machines. (The Verge)

For sizing, use published numbers. NVIDIA’s own recipe table for Nemotron 3.5 Bolt 74B-A9B maps RTX Spark 64GB to its 28GB llama.cpp recipe (65,536-token context) and RTX Spark 128GB to the 32GB recipe (262,144 tokens). It lists no Spark tier below 64GB. Qwen3.8-27B is 55.56 GB in BF16, but Unsloth’s Q4_K_M GGUF is 17.1 GB, so a quantized build is the realistic option on the smaller machines. For very large MoE models, see what it took to run Muse Spark Open locally.

What happened when reviewers ran local agents

The first week of testing was mixed. Tom Warren of The Verge said local AI “feels like a beta aimed at developers right now.”

• Sean Hollister had Hermes, running Qwen3.8-27B, build a daily news briefing. It took 10 minutes and 10% of a 99.9 Wh battery. Claude Opus 5.5 in the cloud took half as long and gave better-sourced results. After a clean reinstall, the same prompt took longer and used 60% of the battery.
• Antonio Di Benedetto used 30% of a Lenovo Yoga 9n’s battery in 36 minutes getting Qwen 3.8 27B to resize five PNGs. The first attempt cropped them wrong.
• Warren said the real strength is leaving an agent running at full throttle overnight, since many tasks were faster on a free cloud model.

On setup, Windows’ new Get Started screen offers installs for Hermes, OpenClaw and Perplexity. Agents install and update through the Windows Store. Warren had to uninstall and reinstall OpenClaw before its local gateway worked.

Tools worth knowing on this platform

• MXC (Microsoft Execution Containers) is now generally available as an OS-level sandbox for agents. NVIDIA says OpenClaw and GitHub Copilot support it now, and Hermes is due by the end of October.
• NVIDIA PAIR is out of beta and adds llama.cpp support. It’s an open source tool that spreads model workloads across PCs on your local network, and it runs on Windows, Linux and macOS.
• NVIDIA Switchyard is an open source library that routes queries between local and cloud LLMs. Hermes plans to add it later in October.

Bottom line: If local inference is the reason to buy, the 64GB and 128GB configs are the ones NVIDIA’s own model recipes target, and you should plan to stay plugged in for long runs. On 24GB and 32GB machines, expect the bundled agents to push you toward cloud or smaller quantized models.