What this laptop is
NVIDIA says RTX Spark laptops are available from ASUS, Dell, HP, Lenovo, Microsoft and MSI. The HP OmniBook Ultra 16 pairs a 20-core Arm CPU with a 6,144-core Blackwell RTX GPU, sharing 128GB of LPDDR5x unified memory at 9400 MT/s. The memory is soldered on the package, so it can't be upgraded later. You have to pick your capacity at purchase.
The rest of the reviewed hardware:
- Display: a 16-inch 2880 × 1800 OLED touchscreen at up to 120Hz. The review lists 100% DCI-P3 coverage and about 500 nits in SDR, rising to 1,100 nits in HDR. The reviewer also notes that the glossy glass picks up reflections.
- Ports: two 40Gbps USB-C ports with DisplayPort 2.1, two 10Gbps USB-A ports, HDMI 2.1, a 3.5mm combo jack and a microSD reader. There is no full-size SD slot, which the reviewer calls an odd choice for a machine aimed at camera work.
- Power and body: a 99Wh battery and a 140W GaN charger. StorageReview measured the chassis at 14.01 x 10.1 inches, tapering from 15.7mm at the rear to 11.6mm at the front, and weighing 4.89 pounds.
- Other features: a 5MP IR webcam with a physical shutter, a discrete TPM, quad speakers tuned by Poly Studio, Wi-Fi 7 and Bluetooth 6.0. The review says Bluetooth runs as 5.4 until Windows support catches up.
What the reviewer measured
The headline use case is local AI work. Techaeris ran LM Studio, Ollama and Jan. It loaded a Llama-3.3-70B-Instruct GGUF at Q4_K_S quantization, which it says takes about 40GB of memory. It reports:
- prompt evaluation of 55–65 tokens per second;
- generation of roughly 14–17 tokens per second;
- generation above 45 tokens per second on 8B and 14B models.
One caveat matters here. The quick take and introduction talk about running 120B models, but the review gives detailed speeds only for the 70B model. I found no measured 120B result in the review, so treat that claim as a capacity argument and not a benchmark.
The reviewer also reports smooth multi-stream 4K and 8K HEVC playback in DaVinci Resolve and Premiere Pro, and quick viewport previews in Blender and Cinema 4D. The review doesn't include project files or a test method, so these are impressions. The solid point is capacity: with a 128GB shared pool, big models and scenes have more room than on a laptop with 12GB or 16GB of dedicated VRAM. That doesn't mean shared memory behaves exactly like dedicated VRAM in every application.
Heat, noise and battery
The trade-offs are the ones you'd expect from a thin chassis with this much silicon:
- Noise and heat: long inference runs bring the fans up to an audible rush and warm the keyboard deck.
- Light use: the reviewer got eight to nine hours from word processing, email, light coding and browsing. HP rates the battery at about 17 hours of mixed light use, so the result is roughly half that.
- Heavy compute: the battery lasts about two to two and a half hours under GPU load.
- Charging: the bundled charger reaches about 50% in 30 minutes and full in about 90 minutes.
Another reviewer saw a more mixed picture. Tom's Guide says the battery life doesn't match the competition and that the laptop demands a steep price. Its verdict is that the machine is for professionals. Notebookcheck's headline advises buyers to choose their RAM configuration carefully, and it notes that the software around the hardware could still use work.
Windows on Arm: check before you buy
For IT admins and developers, compatibility matters more than the spec sheet. Microsoft's documentation says Windows 11 on Arm emulates both x86 and x64 apps. The same page says emulation covers user-mode code only. It doesn't support drivers, and any kernel-mode components must be compiled for Arm64.
That affects the review's remarks about legacy x86 plugins and anti-cheat titles. A user-mode plugin may run under emulation. A driver, a kernel-level security tool or a peripheral utility without an Arm64 build may not run at all. Before you commit, list the essential plugins, device drivers, endpoint security agents, virtualization tools and peripherals, and confirm Arm64 support for each.
NVIDIA and Microsoft say they have prepared developer tooling for the platform. It includes the first PyTorch-CUDA build for Windows on Arm, a native-by-default Python build, a native Conda installer and support for 97% of the top 15,000 Python packages. These are vendor claims, not independent tests. They suggest the developer story is further along than the typical Arm-on-Windows story from a few years ago.
The price confusion
Prices in the review contradict each other, and I can't reconcile them from the review alone:
- Spec table: it says the base model starts at $3,199.99 and the 64GB/2TB build is $4,999.99, and it says HP has published no price for the 128GB custom configuration.
- Price section: the review adds up a configurator total of $5,309 for 128GB/1TB and $6,159 for the tested 128GB/2TB build.
- Notebookcheck: it reports $4,199 for 64GB, a further $1,110 for 128GB and $850 for a 2TB SSD, for a fully configured price of $6,159. That matches Techaeris's arithmetic.
- StorageReview: it lists $6,229.99 for the 128GB/2TB configuration on HP.com, as of October 14, 2026.
- Tom's Guide: it puts the highest-end model, with 128GB of RAM and a 2TB SSD, at $6,299.
The spread runs from about $6,159 to $6,299 for the same 128GB/2TB build. That is what you'd expect from a configurator that changes over time. Check HP's live configurator before ordering.
Who it's for
The review makes a narrow case, and I think the case is reasonable. Developers and ML researchers who need large models on a portable machine, and VFX or 3D artists working with big scenes, get something few laptops offer. The reviewer also argues that running models locally suits NDA-bound or compliance-sensitive work. Be careful with that. Local execution avoids sending prompts and files to a cloud service. It doesn't make your processing compliant. That still depends on your software, data handling and organizational controls.
For ordinary video editing, CAD or office work, the review itself calls 128GB headroom you won't use. It adds that conventional mobile workstations from Dell, Lenovo or Apple cost thousands less. A lower-memory model may suit you better, and you can't add RAM later. The reviewer's praise is real, but the 9.6 score comes from one reviewer looking at a specialized use case. Wait for more independent tests of sustained thermals and Arm64 software support before committing a fleet budget.