I've been waiting for real numbers on the RTX Spark laptop platform, and The Register just delivered: HP's OmniBook Ultra 16, powered by Nvidia's N1X chip, pushes 67.6 tok/s on a 35B model. That's the fastest local AI throughput I've seen in a Windows notebook, and it reframes what a unified-memory laptop is actually good for.
The chip that isn't new
The N1X packs a MediaTek-derived CPU chiplet fused to a Blackwell GPU via Nvidia's NVLink Fusion interconnect. The top SKU — the one tested — runs 20 CPU cores (an even split of Arm X925 and Cortex-A725 big and little cores) and a GPU with 6,144 CUDA cores. That's the same count as the RTX 5070, but with access to far more memory.
The catch is lineage. The N1X is built on the GB10 from the Linux-based DGX Spark, which launched over a year ago. The meaningful change is memory bandwidth: The Register notes the GB10 was marketed at 9,400 MT/s but shipped the DGX Spark at 8,533 MT/s (273 GB/s). The N1X fixes that, landing around 300 GB/s. Higher bandwidth translates directly to faster token generation, so this is the delta that matters for inference.
Where the numbers land
Testing used llama.cpp v0.6.0 across five models (Meta, Google, Alibaba) from roughly 26 to 122B parameters, with a 256-token input and 255-token output per run.
| Model | TTFT (ms) | Generation (tok/s) |
|---|---|---|
| qwen3.6-35b-a3b | 208.20 | 67.64 |
| gemma-4-26b-a4b-it | 201.71 | 55.41 |
| qwen3.5-122b-a10b | 650.45 | 26.23 |
| qwen3.8-27b | 484.91 | 15.76 |
| muse-glimmer-30b | 431.21 | 15.69 |
The 122B model is the headline. 26.23 tok/s on a 122B MoE in a 4.89 lb laptop is the kind of headroom that lets you run a genuinely capable local model with room for a large context window.
Pitted against other hardware on qwen3.6-35b-a3b:
| System | TTFT (ms) | Generation (tok/s) |
|---|---|---|
| HP OmniBook Ultra (N1X) | 208.20 | 67.64 |
| Nvidia DGX Spark | 191.77 | 63.77 |
| HP Z2 Mini G1A (Strix Halo) | 285.30 | 50.37 |
Two caveats from The Register. The DGX Spark and the Strix Halo box are mini PCs with better cooling, not ideal comparison points, which explains the DGX Spark's lower TTFT (more raw compute headroom under thermals). And AMD Strix Halo still wins on value, consistent with what I've seen in other testing.
Where the memory ceiling bites
This is the part I'd weigh hardest. The OmniBook Ultra ships with a 50:50 split — 64 GB to the CPU and 64 GB to the GPU out of the 128 GB pool. On Linux, the DGX Spark allocates memory to the GPU dynamically; Windows doesn't do that yet. You can reconfigure the split, and llama.cpp can overflow into CPU memory with a far smaller penalty than a discrete GPU would incur (because the memory is shared), but a static 50:50 is not ideal for a workload that wants 64 GB+ of GPU-side video memory plus context space.
And fitting the weights is only half the job. Agentic harnesses like Claude Code want large context — The Register calls 128,000 tokens the current sweet spot — which eats a large share of memory before you've even loaded a 200B model (and barely any 200B models exist to load).
What it's not
Price is the biggest asterisk. These start around $2,600 and climb well past $6,000 at top trims. The Register is blunt: RTX Spark isn't a good deal. The hardware is a 2880x1800 OLED, milled aluminum, 4.89 lb, with USB4, HDMI, and MicroSD (no full-size SD or CFexpress).
One honest upside: because it runs CUDA, it's the option if you want Nvidia's ecosystem rather than AMD's ROCm, and you'd prefer Windows over Apple's walled garden. For AI enthusiasts, researchers, and creative pros who need big local video memory, it has a real niche. For the rest, a $2,200 Ryzen AI Max+ 395 mini PC covers most of the same ground cheaper — though at lower bandwidth.
I would not buy this for pure inference value. I would consider it if you specifically want a portable, CUDA-based, 128 GB local model machine and the price doesn't stop you.
Sources: