NVIDIA RTX Spark Laptop Review: StorageReview Clocks 35B Model at 89.5 tok/s

StorageReview ran a 35-billion-parameter coding model at 89.5 tokens per second on a 128 GB NVIDIA RTX Spark laptop, HP's $6,229.99 OmniBook Ultra 16, a model tier that none of the 32 GB or 64 GB RTX Spark laptops it tested could load.

Memory is what decides the buyer's choice, and it sets the shape of this review. The chips desk has already covered the N1X chip itself, the launch, the shipping lineup, and the gaming and power results on the ASUS ProArt P16 and the Surface Laptop Ultra. This piece looks only at local language models: which ones fit at 32, 64 and 128 GB, how fast independent labs measured them, and what each tier costs. Voltgrid did not test these laptops. Every figure below belongs to the outlet that measured it, on the machine it names, and figures from different machines or models are never merged.

The evidence comes from three reviews. StorageReview ran its scripted Windows suite on a 64 GB Dell XPS 16 Creator Edition (Oct. 29) and on two HP OmniBook Ultra 16 units with 32 GB and 128 GB (Oct. 16). Tom's Hardware ran an agentic benchmark on a 128 GB Microsoft Surface Laptop Ultra (Oct. 15). All four machines use the top N1X, with a 20-core Grace CPU and a 6,144-core Blackwell GPU.

What's New

RTX Spark puts NVIDIA's own CPU-and-GPU superchip into a Windows 11 laptop, with CUDA running natively. NVIDIA's spec page lists two laptop versions. The top one pairs the 20-core CPU and 6,144-core GPU with up to 128 GB of LPDDR5X. The lower one has 18 CPU cores, 5,120 GPU cores and up to 64 GB. NVIDIA names CUDA, TensorRT and its NVFP4 number format as the developer stack, which NVIDIA says is the same stack its data-center GPUs use.

The memory is unified, but on Windows it is not shared on demand. The one idea a buyer needs is the split. The CPU and GPU draw from one pool of LPDDR5X. Windows, however, needs the GPU's share fixed in advance, so a model can use only the slice set aside for the GPU. StorageReview reports that Windows does not yet pool CPU and GPU memory the way the Linux software on NVIDIA's DGX Spark does, and that it understands NVIDIA and Microsoft are working on it. In practice, the GPU share is what sets the largest model a laptop can load. That share depends on how much memory the laptop was bought with, because StorageReview notes the memory is soldered.

The price now climbs with memory, and the gap is wide. Dell sells the XPS 16 with RTX Spark at $3,799.99 with the 18-core chip and 32 GB, $3,999.99 with the 20-core chip and 32 GB, $4,799.99 with 64 GB, and $6,999.99 with 128 GB and a 2 TB Gen5 drive, StorageReview reported. HP's OmniBook Ultra 16 starts at $3,199.99 with 32 GB. HP lists a 64 GB, 2 TB model at $4,999.99, and priced the 128 GB build at $6,229.99 as of Oct. 14. Tom's Hardware lists its Surface Laptop Ultra at $5,899 with 1 TB.

NVIDIA's headline claims are 1 petaflop of FP4 AI performance and models of up to 120 billion parameters with a 1-million-token context. NVIDIA's page does not say whether the petaflop figure assumes sparsity, and lists no memory bandwidth. Tom's Hardware puts the platform's shared bandwidth at a maximum of 300 GB/s. StorageReview measured 120 to 122 GB/s in STREAM Triad from the CPU side on every N1X laptop it ran, 2.5 times the 48,298 MB/s of a Razer Blade 16.

What It Solves

At 32 GB, it solves nothing for local language models. StorageReview's 32 GB OmniBook ran with 12 GB given to the GPU and 20 GB to Windows. Its model-load and coding-agent tests need at least 16 GB of GPU memory, so the laptop qualified for none of them. StorageReview's Dell review says the 32 GB XPS configurations get the same 12 GB GPU share. The speed is still there on small models. On MLPerf Client, which runs llama.cpp on CUDA, the 32 GB OmniBook generated Llama 3.1 8B at 56.4 tokens per second, against 56.5 on the 128 GB unit. Memory capacity changes what fits, not how fast a model that fits runs.

At 64 GB with a 32/32 split, 20B and 27B models run, and the 35B tier does not. StorageReview's 64 GB XPS 16 was split 32/32 and measured 30.5 GB usable by the GPU. In its Codex CLI coding-agent test, with gpt-oss:20b served by Ollama at a 64K context, the XPS generated 58.5 tokens per second and passed three of four tasks. Through llama-server, it fit eight concurrent users on Qwen3.8 27B with 8,192-token prompts and delivered 31.9 tokens per second in aggregate. MLPerf Client ran Llama 3.1 8B at 55.7 tokens per second and Phi-4-mini at 96.2. StorageReview says its 35B coding-agent tier needs about 40 GB of GPU memory, so it was out of reach for every 64 GB system it tested.

At 128 GB, a 35B coding model runs at interactive speed. StorageReview's 128 GB OmniBook was split 64/64 and measured 62.5 GB usable by the GPU. It ran Qwen3.6-35B-A3B, a mixture-of-experts model, through the Codex CLI task at a 64K context. It generated 89.5 tokens per second and finished in 72.0 seconds, with three of three runs passing, the quickest of the laptops StorageReview compared. On gpt-oss:20b, the same laptop generated 58.0 tokens per second, level with the 64 GB XPS. Tom's Hardware ran a model of the same name on its 128 GB Surface, in a different test: Qwen 3.6 35B A3B in Unsloth's Q4_K_M quantization, with multi-token prediction, on llama.cpp built from source. Its AgentPerf-local test, built by Artificial Analysis, replays recorded agent traces and reports time to finish. The Surface finished in 8 minutes.

StorageReview does not state the quantization of the models it served through Ollama and llama-server. Neither outlet measured a 120-billion-parameter model, so NVIDIA's 120B claim remains NVIDIA's.

Why Bandwidth Sets Small-Model Speed

Generating each token means reading the model's weights from memory, so a model that fits is limited by how fast memory can deliver them, not by how much memory there is. StorageReview measured Llama 3.1 8B at 56.4 tokens per second on a 32 GB OmniBook and 56.5 on the 128 GB unit, and says token generation at that size is bound by memory bandwidth.

Every RTX Spark laptop shares one design, LPDDR5X-9400 on a 256-bit bus, which Tom's Hardware puts at a maximum of 300 GB/s; NVIDIA's page lists no bandwidth. Capacity decides which models load. Bandwidth decides how quickly the ones that load answer, so extra memory buys larger models, not faster small ones.

Pros and Cons

The strongest case for the platform is the 128 GB result. StorageReview's 89.5 tokens per second on Qwen3.6-35B-A3B came with a passing coding agent on every run. Tom's Hardware found the Surface fastest of the three machines it compared, on both the dense and the mixture-of-experts model. StorageReview's MLPerf numbers show the 32 GB and 128 GB OmniBooks within 0.1 tokens per second of each other on Llama 3.1 8B; StorageReview says token generation at that size is bound by memory bandwidth, which every N1X laptop shares. NVIDIA's spec page is the source for native CUDA. The STREAM Triad comparison with the Razer Blade 16 is StorageReview's.

The weaknesses are all about memory and money. The 12 GB GPU share on 32 GB laptops and the 30.5 GB ceiling on the 64 GB XPS are StorageReview's measurements, as is its note that memory is soldered and the split has to be set ahead of time. The 128 GB prices are Dell's and HP's list prices as StorageReview reported them. StorageReview also measured that a Razer Blade 16 with an RTX 5090 runs 2.3 times faster than the 128 GB OmniBook on models that fit in its 24 GB of graphics memory, though it cannot load the 35B tier at all. Running a coding agent beside other work has a cost too. On the OmniBook, the agent cut the Speedometer browser score from 36.7 to 21.2, StorageReview measured, against a 5% drop on the Blade, whose model sits on a separate GPU.

Who It's For

A developer who wants a 35B coding agent offline. This is the 128 GB buyer. StorageReview's 89.5 tokens per second on Qwen3.6-35B-A3B and three-of-three passes make the case. The price of entry is $6,229.99 for the HP or $6,999.99 for the Dell, and StorageReview names the $6,999.99 XPS as the one for AI developers to buy.

Someone running 20B to 27B models, or a small shared assistant. The 64 GB tier fits. The XPS 16 at $4,799.99 ran gpt-oss:20b at 58.5 tokens per second and served eight users on Qwen3.8 27B at 31.9 tokens per second in aggregate, StorageReview measured. Its RAG test generated answers at 95.0 tokens per second. The $2,200 step to Dell's 128 GB model buys the 35B tier and the Gen5 drive, StorageReview notes.

Not for: anyone buying a 32 GB RTX Spark laptop to run language models locally. At $3,199.99 for the HP or $3,799.99 to $3,999.99 for the Dell, it keeps the GPU speed but loads none of StorageReview's 20B-and-up model tests. A buyer whose models fit in 24 GB and who wants the most tokens per second will get more from an RTX 5090 laptop. Tom's Hardware adds a wider caution: at launch prices, it argues, these machines still buy a lot of cloud tokens, so the case rests on keeping data local.

How It Compares

The three closest machines in the Local AI catalog are NVIDIA's own DGX Spark, AMD's Ryzen AI Halo and Apple's Mac Studio with M5 Ultra. Only one measurement sets RTX Spark against two of them under the same conditions: Tom's Hardware's AgentPerf-local run on Qwen 3.6 35B A3B in Q4_K_M with llama.cpp. On that run, the RTX Spark Surface finished in 8 minutes, a 14-inch MacBook Pro with M5 Max in 10, and the Ryzen AI Halo in 13. On the dense Qwen 3.8 27B, the MacBook Pro took about 25% longer than the Surface, and the Ryzen AI Halo about twice as long, at about 40 minutes. Tom's Hardware notes the Surface uses a 140 W adapter against the 14-inch MacBook Pro's 96 W. The devices also used different speculative decoding on the 27B run: DFlash2 on the RTX Spark and the MacBook Pro, and the model's built-in multi-token prediction on the Ryzen AI Halo.

AMD Ryzen AI Halo. It has the same 128 GB, at 256 GB/s, and was priced at $3,999 with a 2 TB drive at its July launch, StorageReview reported, against $5,899 to $6,999.99 for the 128 GB RTX Spark laptops here. It is a desktop, and it took 13 minutes to the Surface's 8 on Tom's Hardware's mixture-of-experts run.

NVIDIA DGX Spark. It is the same class of chip in a Linux desktop, with 128 GB at 273 GB/s, listed at $6,950 for 128 GB and 4 TB on NVIDIA's marketplace, which showed an Add To Cart button on Oct. 31. As Voltgrid reported, 64 GB partner versions start at $4,999. Its practical edge for model work is the one StorageReview describes: on Linux, CPU and GPU share memory without a fixed split. Tom's Hardware measured it at about 20 tokens per second decoding Qwen 3.8 27B with multi-token prediction. That is a different metric from the laptop's AgentPerf minutes, so the two cannot be ranked against each other.

Apple Mac Studio (M5 Ultra). It offers up to 512 GB of unified memory at 1.2 TB/s, but the 512 GB configuration costs at least $22,299. Tom's Hardware tested the M5 Max MacBook Pro, not the Studio, so no like-for-like Studio result exists yet. In the measurements available, RTX Spark's advantage over the M5 Max MacBook Pro and the Ryzen AI Halo that Tom's Hardware tested is speed on the 27B to 35B models that fit in 128 GB. The Mac Studio's advantage is room for far larger models, at a much higher price.

The full spec table under this story comes from the catalog. Read it with one caution: the bandwidth column compares paper figures, and RTX Spark's 300 GB/s is a maximum reported by Tom's Hardware. The 1.2 TB/s in the table is the Mac Studio's; Tom's Hardware tested a MacBook Pro, and the Surface finished ahead of it in that run.

What to Watch

November 2026: NVIDIA said on Oct. 7 that compact RTX Spark desktops, rated at 140 W in its spec table, would go on sale in November. No independent lab has measured one yet, so whether the higher power rating moves language-model results beyond the 45-80 W laptops is unknown.

Battery results: Both StorageReview reviews say battery life will follow once its new test harness is ready. How long a 128 GB laptop runs a 35B model unplugged is the open question for anyone buying it to work away from a desk.

A 120B measurement: NVIDIA's claim of models up to 120 billion parameters has not been tested by StorageReview or Tom's Hardware. The first recognized lab to publish a gpt-oss-120b result on a 128 GB laptop, with its quantization and GPU split, will show whether the 62.5 GB GPU share StorageReview used is enough, or whether buyers must give the GPU more.

Windows memory sharing: StorageReview understands that NVIDIA and Microsoft are working on unified memory support in Windows. If it arrives, the fixed split stops deciding which models fit, and a 64 GB laptop could reach models that today need 128 GB.

Supply: Tom's Hardware reported lead times for the Surface Laptop Ultra already running to late December.