NVIDIA RTX Spark Laptops: What 128 GB of Unified Memory Means for Local LLMs

NVIDIA's RTX Spark laptops went on sale on 16 October 2026 with up to 128 GB of unified memory, and NVIDIA says compact desktops follow in November. For local AI, the memory matters more than the chip's name: a model that needs 46 GB cannot load on any ordinary laptop graphics card, but it can load here. Two checks come before buying one: how much memory the exact configuration has, and whether the apps you use have a version for Windows on ARM.

What NVIDIA announced

NVIDIA and Microsoft introduced RTX Spark at a Microsoft event on 7 October 2026. NVIDIA's announcement says:

• Laptop preorders opened that day, with availability from 16 October. Compact desktops go on sale in November.
• Computers are coming from Acer, ASUS, Dell, HP, Lenovo, Microsoft, MSI and Gigabyte. Microsoft's is the Surface Laptop Ultra.
• The chip combines an NVIDIA graphics processor and an NVIDIA Grace processor that share one pool of memory.

NVIDIA's announcement gives no prices. Check the maker's page for the configuration you are considering.

The two configurations NVIDIA lists

NVIDIA's page lists the compact desktop with the higher configuration at 140 W.

"Up to" matters. A laptop sold as RTX Spark can have much less than 128 GB, and unified memory in a laptop cannot normally be added later. The memory figure on the product page is the one that decides which models you can run.

Why unified memory matters for local models

A model has to be loaded into memory the graphics processor can reach. On a typical Windows laptop that is dedicated graphics memory (VRAM), often 8 GB to 16 GB, and it sets the ceiling on model size. With unified memory, the processor and graphics share one pool, so a far larger model can load. Apple silicon Macs work the same way, as do some AMD computers; see our 128 GB AMD Strix Halo hardware profile.

The shared pool is not all available to the model. Windows and your other apps use the same memory, and the model also needs conversation memory (KV cache) and room for the model runner. Unified memory for local LLMs explains how to budget it.

What fits

NVIDIA's product page says RTX Spark can run models of up to 120 billion parameters with a 1 million-token context. Its announcement names Qwen 3.8 Flash Next, which it describes as a 125B model, as an example. Those are NVIDIA's statements, and they describe the 128 GB configuration.

Published download sizes give a more useful picture. The sums below are arithmetic from those sizes; Local LLM Guide has not run these models on an RTX Spark computer.

"Left" is what remains for Windows, your other apps, the model runner and conversation memory. On the 64 GB configuration, a 46 GB model leaves little room for a long context. A 1 million-token context needs a large amount of conversation memory on top of the model, so treat that figure as a ceiling for the largest configuration, not a setting to start with. Context windows explained covers how context length uses memory.

Check that your apps run on Windows on ARM

RTX Spark's processor is an Arm design, so these computers run Windows on ARM. Most Windows laptops use Intel or AMD processors, and an app built only for those may run slowly through Windows's translation layer or be unable to use the graphics processor. For local AI, the app needs a Windows on ARM version that supports CUDA, NVIDIA's software for running models on its graphics processors.

Recent release notes show that support arriving:

• Ollama 0.40.16 (9 November 2026) lists CUDA 13.5 support for Windows on ARM.
• LM Studio 0.4.26 (13 October 2026) added CUDA 13 support on Windows ARM and fixed a fault where CUDA was not detected there.
• Open WebUI says it is an early access partner for RTX Spark. Open WebUI is a browser workspace; it still needs a model runner such as Ollama underneath.

Before you buy, check the download page of each app you depend on for a Windows on ARM version. After you buy, confirm the model is running on the graphics processor and not the processor alone. In Ollama, ollama ps shows where a loaded model is running. If it reports the processor only, see GPU not detected in Ollama or LM Studio.

What is not known yet

• Speed. NVIDIA rates the chip at one petaflop of AI performance. That is a measure of raw calculation, not of how many words per second a particular model produces. Local LLM Guide has no measured speeds for any model on RTX Spark. A model fitting in memory does not tell you how fast it answers.
• Battery life and heat under load. Running a large model keeps the chip busy. NVIDIA's announcement gives no figures for battery life while a model is generating.
• Prices for each memory size. These come from each maker, not from NVIDIA's announcement.

How it compares with what you may already have

If you are choosing hardware mainly for local models, choosing a GPU for local AI and what local LLM can my computer run? set out how to match memory to the models you want.

Local does not mean every feature stays local

A model running on an RTX Spark computer processes your prompts on that computer. NVIDIA also describes sending work to cloud services when you need more capacity. As with any setup, check where each app sends prompts and files. The local AI privacy guide lists what to look for.

Related reading

• Local LLM hardware requirements: RAM and VRAM
• RAM and VRAM calculator
• Run local AI on Windows: LM Studio and Ollama setup

Sources

• NVIDIA's announcement of RTX Spark with Microsoft, 7 October 2026
• NVIDIA's RTX Spark product page and specification table
• Liliputing's report on NVIDIA's Arm-based chips for Windows laptops
• Ollama 0.40.16 release notes
• LM Studio changelog
• Open WebUI's note on RTX Spark