NVIDIA RTX Spark Laptops for Local AI: 128GB, 300GB/s and What Fits

NVIDIA’s RTX Spark laptops went on sale on 16 October 2026: Windows laptops built around the same chip as the DGX Spark, with up to 128GB of memory shared between the processor and the graphics. For local AI that is the headline. A laptop that can hold a 120-billion-parameter model has, until now, meant a MacBook Pro or an AMD Strix Halo machine.

This guide covers what the chip is, how much of the memory the GPU can actually use, what fits at each memory size, what the laptops cost, and which local-AI software already supports them. BFWR has not tested an RTX Spark laptop. Specifications are NVIDIA’s and Microsoft’s, prices are read from Microsoft’s store, the fits and speeds are estimates from our own hardware database, and the measured results are credited to the people who measured them.

Quick answer

• What it is: the N1X chip, NVIDIA’s GB10 from the DGX Spark, in Windows laptops. 6,144 CUDA cores and a 20-core Arm CPU, or 5,120 cores and 18 for the cheaper version.
• Memory: 24GB to 128GB of LPDDR5X on a 256-bit bus. NVIDIA’s porting guide gives 300 GB/s, slightly more than the DGX Spark’s 273 GB/s.
• Usable by the GPU: not all of it. On a 128GB machine, Windows lets the GPU address about 110GB.
• What fits: 28 of the 30 models in our directory at 128GB, including gpt-oss-120b, the same as a DGX Spark, Strix Halo or M5 Max machine. At 24GB, 14 of 30.
• Price: the Surface Laptop Ultra runs from $2,599.99 (24GB) to $5,899.99 (128GB).
• Software: llama.cpp, Ollama, LM Studio and ComfyUI all ship Windows-on-Arm builds with NVIDIA GPU support.

What RTX Spark is

NVIDIA’s specification page lists three versions of the chip, called N1X:

NVIDIA’s own pages disagree on the lower chip’s limit: the specification page says up to 32GB and the product page up to 64GB. The laptops on sale pair it with 24GB or 32GB, and every configuration with 48GB or more uses the top chip. The petaflop figure is for 4-bit (FP4) maths with sparsity, as Microsoft’s footnote spells out, so it is not comparable with ordinary GPU specifications.

The chip is the GB10 from the DGX Spark. ServeTheHome calls the N1X “the consumer-facing name of our tried-and-true friend in the Linux mini-PC market: the NVIDIA GB10.” The differences are faster memory (LPDDR5X-9400 against the DGX Spark’s 8533), a much lower power limit in laptops (45 to 80 W against 140 W), Windows on Arm instead of Linux, and no high-speed network card. Our DGX Spark vs Strix Halo comparison covers the desktop version.

300 GB/s, not 600

Some early coverage gave 600 GB/s for memory bandwidth. NVIDIA’s developer porting guide settles it: “The interface is 256 bits wide, with a bandwidth of 300 GB/s.” The 600 GB/s figure, per ServeTheHome and StorageReview, is the NVLink-C2C connection between the CPU and GPU halves of the chip, not the memory.

Bandwidth sets how fast a model generates text, because every token reads the active weights from memory. At 300 GB/s, RTX Spark sits between the AMD Strix Halo (256 GB/s) and Apple’s M5 Max (614 GB/s). Our tokens-per-second guide explains the arithmetic.

How much memory the GPU can use

The memory is shared, but the GPU does not get all of it. NVIDIA’s porting guide says the memory is split into three regions: a block reserved for the GPU, which Windows reports as dedicated GPU memory; shared memory both can use; and “CPU-only system memory”, which “cannot be used for GPU allocations”. It gives the rule for the shared part: “The nominal shared system memory size is the post-carveout capacity minus 16 GB. That value is clamped to a minimum of 50% and a maximum of 80% of the post-carveout capacity.”

NVIDIA does not publish the size of the reserved block. Using the default sizes ServeTheHome reported from NVIDIA’s briefing and applying NVIDIA’s formula, we calculate the GPU’s budget as:

The 128GB figure checks out against a real machine. StorageReview read a Task Manager screenshot of a 128GB Surface Laptop Ultra showing “110GB of total GPU memory”, and the formula reproduces it. Microsoft’s own footnote agrees in principle: the amount “addressable by the GPU depends on system configuration and workload and is less than the total.” NVIDIA’s guide also warns that “Allocating the full GPU budget can leave too little host memory and can make the system unresponsive”, so leave headroom.

What fits at each memory size

We added the four RTX Spark configurations to the BFWR hardware database, using the GPU budgets above. These counts come from the same calculation as the device pages, at 4-bit (Q4_K_M):

Two points stand out. The 24GB laptop holds exactly as many models as a 16GB graphics card, because a third of its memory is held back from the GPU. And at 128GB, capacity is a tie: a DGX Spark, a 128GB Strix Halo machine and a 128GB M5 Max Mac Studio all fit the same 28 models in our directory. What differs is speed, and that comes down to bandwidth.

Estimated generation speeds on the 128GB laptop, from bandwidth (BFWR estimates, not measurements):

• gpt-oss-20b: about 82 tokens/s
• gpt-oss-120b: about 69 tokens/s
• Qwen3.6 35B-A3B: about 108 tokens/s
• Qwen3.8 27B: about 12 tokens/s
• Llama 3.3 70B: about 5 tokens/s

Mixture-of-experts models such as gpt-oss and Qwen3.6 35B-A3B read only a few billion parameters per token, so they are far faster than dense models of a similar size. On a 300 GB/s machine they are the models to choose.

Measured results so far

Two credible sources have published results for the laptop:

• Artificial Analysis ran an agent workload of 168 turns on a prototype 128GB Surface Laptop Ultra with llama.cpp. On Qwen3.8 27B, “the RTX Spark laptop completes the workload in 18.7 minutes, close to DGX Spark (17.3), 2.0x faster than the MacBook Pro (37.6)”, where the MacBook Pro was an M5 Pro with 64GB. On Qwen3.6 35B-A3B it took 7.6 minutes against the DGX Spark’s 7.3. Its own caveat: “the runtime, speculative decoding and 4-bit weight format differ between systems.”
• ServeTheHome‘s review found the laptop slightly behind the DGX Spark under concurrent load, “which we would expect because it has a higher power budget”, and that “decode performance tends to favor the M5 Max 128GB due to the memory bandwidth.” It also found strong prompt processing. The reviewers estimate battery life under heavy agent workloads at “closer to an hour”.

For the DGX Spark itself, the llama.cpp maintainers’ benchmark gives gpt-oss-120b at 58.72 tokens/s for generation. The laptop has 10% more bandwidth but less than half the power, so treat that as a ceiling rather than a laptop figure.

What the laptops cost

NVIDIA’s announcement lists models from ASUS, Dell, HP, Lenovo, Microsoft and MSI, “starting at $2,599”. Microsoft’s US store, read on 24 October, prices the Surface Laptop Ultra at:

For comparison, ServeTheHome reported the 128GB DGX Spark at $6,950 on 3 October, and Framework’s 128GB Strix Halo desktop board was $3,449 before storage on 12 October; both are desktops. Microsoft’s Surface RTX Spark Dev Box desktop is $5,999.99 on pre-order, and NVIDIA says RTX Spark desktops arrive in November.

Which configuration for local AI

• Skip 24GB for local AI. About 16GB reaches the GPU, which is the same model range as a 16GB graphics card, at a laptop price. NVIDIA’s own “Help me choose” page says the 24GB and 32GB tiers have “Agents run in the cloud”.
• 64GB is the sensible minimum for running large models: about 48GB for the GPU, enough for 70B dense models at 4-bit or Qwen3.6 35B-A3B with plenty of context.
• 128GB is the point of the platform. It is the only size that holds gpt-oss-120b, and the $1,600 step from 64GB buys roughly 62GB more for the GPU.
• Against a Mac: a 128GB M5 Max machine generates text roughly twice as fast on the same models, because of its 614 GB/s bandwidth. RTX Spark’s case is CUDA, Windows and stronger prompt processing, which matters for long documents and agent tools.

Software support

RTX Spark runs Windows 11 on Arm, so software needs native Arm builds with NVIDIA GPU support. As of 24 October:

• llama.cpp publishes official Windows Arm64 CUDA 13.4 builds with every release. The pull request that added them notes validation “on a native ARM64 RTX Spark system”.
• Ollama added “CUDA on Windows ARM64” in v0.32.3. Its documentation does not yet mention RTX Spark by name.
• LM Studio 0.4.26 “Added CUDA 13 support on Windows ARM.”
• ComfyUI has shipped an Arm64 NVIDIA portable build since v0.40.0.
• PyTorch with CUDA for Windows on Arm comes from NVIDIA’s own package index, not pytorch.org, which offers only CPU builds for Windows on Arm.

Older Windows tools without Arm builds run through Microsoft’s x86 emulation, but GPU-accelerated AI software generally needs a native build. Check before you buy if you rely on a specific tool.

FAQ

Is RTX Spark the same as DGX Spark?

Same chip, GB10, in a different form: a Windows laptop with faster memory (300 against 273 GB/s) but a 45 to 80 W power limit instead of 140 W, and no high-speed network card.

Can an RTX Spark laptop run gpt-oss-120b?

The 128GB model can. Its 65GB of 4-bit weights fit in the roughly 110GB the GPU can address. We estimate about 69 tokens/s from bandwidth; measured results on the desktop DGX Spark are about 59.

How much of the 128GB can the GPU use?

About 110GB. NVIDIA’s memory split reserves part of the memory for the CPU only, and a real 128GB Surface reported 110GB of total GPU memory.

Does Ollama work on RTX Spark?

Ollama ships a Windows Arm64 build with CUDA support from v0.32.3, though its documentation does not list RTX Spark yet. llama.cpp and LM Studio state Windows on Arm CUDA support directly.

Related guides and tools

• Which models fit each RTX Spark memory size: 24GB, 32GB, 64GB and 128GB
• DGX Spark vs Strix Halo
• Strix Halo vs Apple M5 Max
• Best GPU for AI
• RTX Spark 128GB in the hardware matrix and the GPU comparison

Sources

• NVIDIA: RTX Spark specifications, product page, porting guide system overview and unified memory, and the 16 October launch announcement
• Microsoft Surface Laptop Ultra and Microsoft Store prices
• Artificial Analysis: RTX Spark AgentPerf results
• ServeTheHome: N1X launch overview and Surface Laptop Ultra review
• StorageReview: RTX Spark pre-orders and memory split
• llama.cpp Windows Arm64 CUDA builds, Ollama v0.32.3, LM Studio 0.4.26, ComfyUI v0.40.0
• llama.cpp DGX Spark benchmarks

Updated 24 October 2026. Prices and software support were read on that date. Model fits and tokens-per-second figures are BFWR estimates from the BFWR hardware database; measured results are credited to their sources.