Best Local LLM for RTX Spark (2026): Set the GPU to 96GB

Bottom Line

• 128GB RTX Spark: run gpt-oss 120B. MindStudio measured ~65 tok/s in llama.cpp. Its MXFP4 weights are 59.03 GiB, so first raise the GPU memory share to 96GB.
• The default split is the trap. The 128GB machines ship split roughly evenly. StorageReview measured 62.5GB GPU-usable at the default. That leaves almost no room for the 120B model’s cache.
• 64GB RTX Spark: run Qwen3.6 35B-A3B. StorageReview measured 89.2 tok/s in Ollama. gpt-oss 120B does not fit the 64GB tier at all.
• Avoid dense 27B-32B models. They measured 11.8-13.4 tok/s. NVIDIA publishes no memory bandwidth; the theoretical figure is about 300 GB/s.
• Use Ollama first. Windows Smart App Control blocked unsigned llama.cpp builds in XDA’s test. Ollama ran on both laptops.

Still choosing a machine? Compare it with the tested hardware list with current prices.

Best Models by Memory Tier

Every speed below was measured by a named reviewer in October 2026. We did not test an RTX Spark machine. Numbers do not transfer between laptops, runtimes or power modes.

The pattern is clear. Mixture-of-experts models with few active parameters run fast. Dense models run slowly. Pick from the top of the table.

The MindStudio numbers come from one article and no second review repeats them. Treat the 65 tok/s figure as one data point.

Set the GPU Memory Split First

This is the step most buyers miss. The pool is shared, and the GPU does not get all of it.

• ASUS ProArt P16: PCWorld says MyAsus “won’t let you devote more than 96GB” to the GPU. The reviewer ran at about 96GB.
• Surface Laptop Ultra: StorageReview measured 62.5GB GPU-usable at the default 64/64 split, adjustable up to 96GB.
• 64GB Lenovo Yoga 9n: XDA saw 55.4GB of total GPU memory in Windows at a 48GB split. gpt-oss 120B needs “roughly 65GB” for its weights alone, so it does not fit.

Why 96GB matters for gpt-oss 120B. The weights are 59.03 GiB (63.4 GB). At the default share, only a few GB remain for the KV cache and buffers. At 96GB, MindStudio loaded it using about 61GB, with room for long context.

The memory is soldered. XDA’s first warning: the 18-core chip “tops out at 64GB, while the full 20-core version goes up to 128GB.” Buy the tier you need on day one.

Why Dense Models Are Slow Here

NVIDIA’s RTX Spark spec page lists memory size and “up to 1 petaflop” of FP4, but no memory bandwidth. Developers Digest points out that the 600 GB/s in the launch post “is the link between the CPU and GPU, not the memory.”

StorageReview computed the memory figure: LPDDR5X at 9,400 MT/s on a 256-bit bus “works out to about 300GB/s of theoretical bandwidth.” That is close to a DGX Spark (273 GB/s) and about half of an M5 Max (614 GB/s).

A dense model reads every weight for every token. A 22GB dense file on ~300 GB/s has a ceiling near 14 tok/s, and MindStudio measured 13.4. An MoE model reads only its active experts, so gpt-oss 120B (about 5B active) runs about five times faster despite having four times as many parameters.

For a decode-heavy workload on a dense model, a 32GB desktop card is faster. For a model that does not fit 32GB, RTX Spark wins. PCWorld ran the same gpt-oss 120B job on a 128GB ProArt P16 and an RTX 5090 desktop: 33 min 13 s vs 52 min 51 s, about 1.6 times faster. The reviewer added: “don’t read it as ‘RTX Spark is faster than the RTX 5090 at AI.’ It isn’t.”

Software: What Runs on Windows on Arm

If you use WSL2, raise the RAM limit in .wslconfig. WSL2 does not see the full pool by default.

RTX Spark vs DGX Spark vs Strix Halo vs Mac

Prices are as reported in October 2026. Check current listings before you buy.

Reported RTX Spark 128GB prices: Surface Laptop Ultra $5,899.99 (StorageReview, 15 October); Lenovo Pro 9n top configuration $6,099.99 (Notebookcheck via WindowsForum, 8 October); MSI Prestige N16 Flip AI+ $6,799.99 (MSI US store via The Desk Brief, 19 October); ASUS ProArt P16 $6,999 (Micro Center, per PCWorld, 29 October). The cheapest RTX Spark laptops start at $2,599.99-2,999.99 with 24-32GB.

Our read. If you need a laptop with CUDA, the 128GB RTX Spark runs gpt-oss 120B on the GPU at a usable speed. If the machine can stay on a desk, a Strix Halo box gives the same 128GB class for about half the price. If decode speed on large dense models matters, the MacBook Pro’s bandwidth is about twice as high.

Sources

• StorageReview: Surface Laptop Ultra for Business review (15 October 2026: MLPerf Client, Ollama tok/s, 62.5GB GPU-usable, ~300 GB/s theoretical, $5,899.99 as tested)
• XDA: 5 things nobody tells you before buying an RTX Spark laptop (25 October 2026: memory tiers, split, Smart App Control, CUDA on Windows on Arm, Linux)
• PCWorld: ASUS ProArt P16 review (29 October 2026: 96GB GPU cap, RTX 5090 comparison, $6,999)
• MindStudio: Running 120B+ LLMs locally on the RTX Spark laptop (17 October 2026: gpt-oss 120B, Qwen3.8 27B and Flash Next, vLLM in WSL; single source)
• Developers Digest: RTX Spark for local LLMs (14 October 2026: no official bandwidth figure)
• NVIDIA RTX Spark product page (memory tiers, 1 petaflop FP4)
• gpt-oss 120B MXFP4 size from the Hugging Face API (59.03 GiB). DGX Spark, Strix Halo and MacBook Pro prices from our price reference, read 2026-10-26 to 2026-10-31

See Also

• RTX Spark Slow on Linux? (2026) — why native Linux runs at ~780 MHz plugged in, and the measured fixes
• Best Models for NVIDIA DGX Spark
• DGX Spark vs Strix Halo
• Best Local LLM for 128GB RAM
• Which Strix Halo Mini PC Should You Buy?
• Is the DGX Spark Worth It?