The advice is simple: if local AI is your reason to buy, skip the 32GB version. If you want an Arm laptop for creative work and gaming, 32GB is fine.
What was tested
The reviewed laptop has these specifications:
- Processor: a 20-core Grace Arm CPU (10 performance cores, 10 efficiency cores) with a Blackwell GPU that has 6,144 CUDA cores.
- Memory: 32GB of LPDDR5X-9400, soldered on and shared between the CPU and GPU.
- Storage and OS: a 1TB PCIe 5.0 SSD and Windows 11 Home.
- Display: a 16-inch, 2880×1800, 120Hz touch OLED rated at 1,100 nits and DisplayHDR True Black 1000.
- Ports: two USB4 Type-C ports, two 10Gbps USB-A ports, HDMI 2.1, microSD and a headset jack.
- Charger: a 140W USB-C GaN adapter. The laptop weighs 4.89 pounds, and HP rates the battery at up to 17 hours.
The 32GB/1TB model starts at $3,200 at Best Buy and HP. The 64GB/2TB and 128GB/2TB versions are sold as configurations directly from HP.
StorageReview tested a 32GB and a 128GB unit side by side. It reports the 32GB, 1TB unit is Best Buy's $3,199.99 configuration, which is also HP's starting price. It lists the 64GB/2TB model at $4,999.99 and the 128GB/2TB build at $6,229.99 as of October 14. Treat these as launch-window prices that may change.
Where 32GB hits the wall
The review says the 32GB machine splits memory by default. The GPU gets 12GB. Windows holds a large share for itself. Most AI apps saw only about 20GB of usable VRAM.
The review's own memory arithmetic is a bit muddled. It mentions a 12GB GPU share, a 10GB shared pool and 10GB reserved for Windows. StorageReview describes the same unit more cleanly. It set the 32GB machine to a 12GB GPU and 20GB CPU split. This article treats the roughly 20GB as the app-visible budget the reviewer saw, not a full memory map.
The Hermes agent test went like this:
- Hermes alone: the app would not offer any Qwen 3.6-35B-A3B variant above IQ2_M, a heavily compressed 2-bit quantization.
- LM Studio: it loaded the IQ4_XS version as a partial offload. Hermes needs 64K context, which made reliable loading hard.
- What finally worked: IQ4_XS with 64K context and one parallel slot. It still failed several times.
- When it ran: responses and reasoning were good and token speed was fast. The reviewer put it on par with GPT-6.1 Sol and slightly slower than Opus 5.5. A prompt might work once and fail the next time.
The reviewer notes that a smaller model would have worked, but he didn't want to settle for a weaker one.
What fits comfortably
Models that fit in memory did well:
- Qwen 3.5 9B: just over 46 tokens per second.
- Qwen 3.5 27B: about 15 tokens per second. Hermes completed every task, at just over 100 seconds per task.
- Backends: CUDA beat Vulkan.
- Image generation: FLUX.2 Klein 4B took 7.6 seconds per image and SDXL took 22.2 seconds.
These are one reviewer's numbers. Quantization, context length, backend and settings all move them.
StorageReview found the same split. Its 32GB unit matched the 128GB unit on MLPerf Llama 3.1 8B, at 56.4 tokens per second against 56.5. But its 12GB GPU share kept it out of every model-load and coding-agent test. That lines up with the How-To Geek verdict: speed isn't the issue, capacity is.
Can't you just change the split?
The review says NVIDIA's tool lets you move the GPU's share between 4GB and 16GB. That doesn't add physical memory, so the models that fit stay about the same.
StorageReview adds that the split is a Windows limitation. It says Windows does not yet support unified memory across the CPU and GPU the way the Linux stack on DGX Spark does. So the allocation has to be set in advance. It says NVIDIA and Microsoft are working on better support, but that is a report, not a shipping feature.
Gaming and creative work
The gaming results in the review were better than expected:
| Game (native 2880×1800) | Extreme | Low |
|---|---|---|
| Forza Horizon 5 | 79 fps | 118 fps |
| Forza Horizon 6 | 45 fps | 74 fps |
The reviewer says Forza Horizon 5 looked CPU- or emulation-bound, with the GPU at about 60% use. Forza Horizon 6 was GPU-bound. The GPU sat near 70°C and held roughly 2.2–2.4GHz without throttling. The same reviewer's Lenovo Yoga Pro 9i with an RTX 5060 also hit 79 fps in Forza Horizon 5 on Extreme.
Other reviewers found a more mixed picture. Tom's Guide's test set, as summarized in the submitted research, had Cyberpunk 2077 at 18fps at maximum settings and native resolution. It rose to 38–47fps at default settings and 58–65fps at 1080p with DLSS. So the Forza results don't carry over to every demanding title. How-To Geek also lists emulation capping frame rates in older x64 games as a con.
Other points from the review:
- CPU: the N1X led its comparison group in multi-core work and beat the Snapdragon X2 Elite Extreme in some tests. The text doesn't include scores.
- Battery vs plugged in: performance stayed about the same on battery, though best-performance mode drains it faster.
- Memory sharing: don't game while a local model is loaded, because the GPU and models share one memory pool.
- Build: the keyboard, haptic trackpad, metal chassis and compact charger got praise. The charger's plug doesn't block neighboring outlets.
Battery life: read the fine print
HP rates the 99Wh battery at up to 17 hours. Tom's Guide, per the research, recorded 10 hours 52 minutes in a continuous web-browsing test. The two figures measure different workloads and aren't directly comparable. HotHardware's pros-and-cons list also includes mediocre battery life.
What the platform promises
NVIDIA's launch announcement says RTX Spark laptops are available globally today from ASUS, Dell, HP, Lenovo, Microsoft and MSI. It cites up to 128GB of unified memory and up to 1 petaflop of FP4 AI compute. It names guided setup for agents including GitHub Copilot, Hermes, OpenClaw, Perplexity, Pi and WorkBuddy. It also names NVIDIA OpenShell and Microsoft Execution Containers as ways to control agent permissions.
Those are "up to" figures for the top configuration. They don't show that the 32GB HP model can run large models. The review's gap between marketing and practice is the point of the whole piece.
Who should buy which configuration
- Creators and gamers: 32GB gives up little on CPU performance, though it costs a lot. StorageReview found the 32GB unit within about 2 percent of the 128GB unit on CPU, MLPerf and RAG work.
- Local AI buyers: get 64GB or 128GB. The reviewer says to avoid the 24GB and 32GB tiers for this use. His call is based on a 35B-class model at 64K context, not a universal minimum.
- Everyone: check your target model, quantization, context length and background apps before paying more for memory.
The review gives it 7/10. The reviewer says it's excellent hardware, but the 32GB model isn't the one for local AI.