RTX Spark Local LLM Guide: N1X Specs, Models for 24–128 GB, RTX Spark vs DGX Spark

For RTX Spark local LLM work, the memory split matters more than the GPU. NVIDIA's N1X pairs 6,144 CUDA cores with 300 GB/s of bandwidth and up to 128 GB of unified memory, but Windows hands the GPU only part of it. A 128 GB machine runs gpt-oss-120b at roughly 36 tokens per second by our estimate; a 32 GB one holds about what a 16 GB graphics card does.

RTX Spark is NVIDIA’s platform for Windows PCs. Its chip, the N1X, joins a Grace CPU and a Blackwell RTX GPU over one pool of memory, much as the GB10 does in DGX Spark. NVIDIA announced it on 31 May 2026, and the first laptops ship from 16 October.

The facts below come from NVIDIA’s product page, newsroom post and Windows-on-Arm porting guide, Microsoft’s store, and the release notes of the tools you would run, all checked on 14 October 2026. Where NVIDIA publishes nothing, we say so. Speeds are estimates from the method behind our tokens-per-second calculator, not benchmarks. For Mac Studio and Ryzen AI Max comparisons, see our desk-side hardware comparison.

RTX Spark N1X: what NVIDIA has published

The bandwidth comes from NVIDIA’s porting guide for developers, which gives 300 GB/s on a 256-bit interface; third-party reports had quoted anything from 273 to 307 GB/s. The product page headlines up to 1 petaflop of FP4 AI performance without a sparsity note. Microsoft’s footnote for its own 120B-model claim says it rests on theoretical FP4 throughput using sparsity, and NVIDIA’s DGX Spark documentation gives GB10 the same 1 petaflop with sparsity. One view of NVIDIA’s specification table also shows 140 W for the 6,144-core chip without naming the form factor.

Some reports call the smaller chip N1. NVIDIA’s product page and Microsoft’s store call both versions N1X. NVIDIA says the platform runs models of up to 120 billion parameters with a 1M-token context.

Laptops, prices and availability

NVIDIA’s product page lists six launch laptops: ASUS ProArt P16, Dell XPS 16, HP OmniBook Ultra 16, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra and MSI Prestige N16 Flip AI+. Its announcement adds that Acer and GIGABYTE will follow and that small desktops are part of the platform; an NVIDIA blog from IFA showed an Acer compact desktop concept.

Microsoft’s US store shows a clear price ladder for the Surface Laptop Ultra, a 15-inch machine:

Every configuration lists a release date of 16 October 2026. NVIDIA publishes no device prices, and the other makers set theirs by configuration and region. Doubling memory from 64 to 128 GB adds $1,600 on Microsoft’s ladder, so decide which models you need before choosing a configuration.

How much memory the RTX Spark GPU can use

NVIDIA’s porting guide splits unified memory into three regions. A dedicated carveout belongs to the GPU and appears in Windows as dedicated GPU memory. Of what remains, a shared region is open to both CPU and GPU, and the rest is reserved for the CPU. The guide sets the nominal shared size at the post-carveout memory minus 16 GB, clamped to between 50% and 80% of that post-carveout memory.

CUDA allocations land in the carveout first and spill into the shared region once it is full. The guide notes that shared allocations use smaller pages, so access can be slower, and that allocating the whole GPU budget can leave Windows too little memory and make the system unresponsive.

Our sizing takes the smaller of two limits: this budget, and the 80% of unified memory that our VRAM calculator always leaves for the operating system. At 128 GB the two agree at about 102 GB, the same figure we use for DGX Spark. From 64 GB down, the Windows rule is the tighter limit, and at 24 and 32 GB it is far tighter. NVIDIA does not publish the carveout size; Windows reports it as dedicated GPU memory.

Which models fit in 24, 32, 48, 64 and 128 GB

File sizes are the GGUF downloads on Hugging Face from ggml-org, LM Studio’s community account and Unsloth, checked 14 October 2026. Context figures use each model’s published attention shape with a 16-bit cache, the same method as our VRAM tier guide.

Two conclusions follow. The 24 and 32 GB machines give the GPU roughly what a 12 or 16 GB graphics card has, so they suit small models; the case for RTX Spark as a model machine starts at 48 GB. And only the 128 GB configuration holds the 120B class that NVIDIA advertises, where gpt-oss-120b fits with room for several long conversations. Larger mixture-of-experts models such as GLM-5.3-Flash fit 128 GB only as 1-bit community builds, as our GLM-5.3 deployment guide explains.

RTX Spark LLM performance: bandwidth-bound estimates

Each generated token reads the model’s active weights from memory once, so bandwidth sets a ceiling on speed. Our calculator assumes a runtime achieves 60% of peak bandwidth. The figures below are for one user, with a 2,048-token prompt, a 512-token reply and a 16-bit cache.

Tokens per second, estimated with the calculator’s method; none of these are measurements. Time to first token is left out because NVIDIA publishes only a headline FP4 figure for N1X, not the dense rates prefill depends on.

Two patterns stand out. Mixture-of-experts models suit this hardware: gpt-oss-120b reads about 5 GB per token where a dense 31B model at 4-bit reads 19 GB, so the larger model runs about four times faster. Concurrency costs speed per person: with four simultaneous requests, gpt-oss-120b falls to about 17 tokens per second each, roughly 68 in total.

To measure your own, llama.cpp’s llama-bench tool reports prompt-processing and generation speed for any model file. Run it before you trust any estimate, ours included.

RTX Spark vs DGX Spark

On paper the two share a GPU class, a CPU layout and a memory ceiling, with RTX Spark’s memory about 10% faster. NVIDIA does not say they are the same silicon, and the differences lie around the chip. DGX Spark runs NVIDIA’s Linux-based DGX OS, so the engines teams use on servers, vLLM among them, run as they would in a data centre, and up to four units link into one larger pool. RTX Spark is a Windows PC: native builds of Ollama, LM Studio and llama.cpp, no clustering ports, and in its 45–80 W rating a power envelope well below GB10’s 140 W, which makes long, sustained runs worth testing.

Prices have moved. NVIDIA’s DGX Spark page prints none. NVIDIA raised the 128 GB Founders Edition from $3,999 to $4,699 in February 2026, and on 2 October announced a 64 GB model from $4,999, sold through partners from 23 October. Guru3D reported the same day that NVIDIA had also raised the 128 GB Founders Edition to $6,950. A 128 GB Surface Laptop Ultra costs $5,899.99.

Choose DGX Spark as a team’s development bench that mirrors NVIDIA servers. Choose RTX Spark when one person wants a large model on a Windows laptop.

RTX Spark vs RTX 5090, Apple M5 and Strix Halo

• RTX 5090, 5080 and 5070. NVIDIA lists 32 GB at 1,792 GB/s for the RTX 5090, 16 GB at 960 GB/s for the 5080, and 12 GB at 672 GB/s for the 5070, which has the same 6,144 CUDA cores as the top N1X. A desktop card streams any model it can hold two to six times faster than RTX Spark. RTX Spark’s advantage is capacity: at 128 GB it holds gpt-oss-120b, which none of those cards holds alone, at a fraction of the 5090’s 575 W. Our GPU buying guide covers the cards in detail.
• Apple M5, M5 Pro and M5 Max. Apple’s MacBook Pro specifications list 153 GB/s for M5 with up to 32 GB, 307 GB/s for M5 Pro with up to 64 GB, and 460 or 614 GB/s for M5 Max with up to 128 GB. M5 Pro matches RTX Spark’s bandwidth but stops at 64 GB; M5 Max doubles it at the same 128 GB ceiling. Under our 80% rule a 64 GB Mac offers about 51 GB to models, slightly more than the 48 GB Windows gives a 64 GB RTX Spark. Our Mac guide for local models has the model picks.
• AMD Strix Halo. AMD’s Ryzen AI Max+ 395 pairs 128 GB of LPDDR5x with 256 GB/s and a 40-compute-unit Radeon GPU: the same capacity at about 15% less bandwidth, on AMD’s software stack rather than CUDA.

For shared serving rather than a personal machine, compare data-centre GPUs in our RTX PRO 6000, H100 and H200 comparison.

Setting up RTX Spark for local models

1. Update Windows and the NVIDIA driver from the laptop maker. NVIDIA added Windows on Arm support to the CUDA Toolkit in version 13.4, announced on 9 September, and llama.cpp’s ARM64 build is compiled against it.

2. Choose a native ARM64 runtime. Ollama’s release notes added CUDA on Windows ARM64 in version 0.32.3 (23 July 2026), and version 0.40.4 (12 October) ships a Windows ARM64 package. LM Studio 0.4.26 (13 October) added CUDA 13 on Windows ARM, and its build 5 fixed CUDA detection on those machines. llama.cpp’s releases include a win-cuda-13.4-arm64 package. NVIDIA’s guide says many x86 programs run under emulation, but native ARM64 builds perform better.

3. Pull a model that fits your tier and confirm it runs on the GPU.

   ollama pull gpt-oss:20b
   ollama run gpt-oss:20b "List three risks in this contract clause: ..."
   ollama ps

   The PROCESSOR column of ollama ps shows how the model splits between GPU and CPU; aim for 100% GPU.

4. Set the context window explicitly. Ollama derives its default from GPU memory, and its documentation recommends at least 64,000 tokens for agents and coding tools. On Windows, set OLLAMA_CONTEXT_LENGTH as a user environment variable, then quit and relaunch Ollama.

5. Benchmark before you standardise. Run llama-bench -m <model>.gguf and compare the result with the estimates above.

6. Treat Linux engines as experimental. vLLM’s documentation says it does not support Windows natively, and NVIDIA’s CUDA on WSL guide does not yet cover Windows on Arm.

If colleagues will call the laptop’s model over the network, read our guide to serving Ollama to several users first: Ollama’s local API has no authentication.

How VDF AI fits

An RTX Spark laptop is a personal machine, and VDF AI’s self-hosted platform is not a laptop workload. Our infrastructure requirements name Linux as the production target and list Windows, through Docker Desktop and WSL 2, for evaluation only.

Where an RTX Spark helps is model evaluation. VDF AI Router can register an Ollama endpoint as a model deployment next to the cloud models your policy approves, so a pilot can try a 120B-class model on one machine. When the workload moves to rack GPUs, applications keep calling the router and only the deployment behind it changes, and air-gap mode keeps routing on local models.