In a comparison published Oct. 26, 2026, Patrick Campanale reported that Hermes offered three one-click models on the 48GB ASUS Zenbook A16 and no compatible models on the 32GB HP OmniBook Ultra 16 with NVIDIA RTX Spark N1X. Campanale also found the RTX Spark faster when a model fit, so the Zenbook’s easier model selection did not make it the faster system in his reported test.
What Hermes offered on each laptop
Campanale tested the HP OmniBook Ultra 16 with 32GB of unified memory and the ASUS Zenbook A16 (2026) with 48GB of RAM. On the Zenbook, Hermes offered Qwen 3.8 27B, Nemotron 3.5 Bolt, and Qwen 3.6 35-A3B for one-click installation. The OmniBook’s Hermes list showed no compatible model, even with its dedicated-VRAM setting at 16GB, the maximum Campanale could set in that configuration.
| Tested system | Memory | Hermes model list | Reported execution path |
| HP OmniBook Ultra 16 with NVIDIA RTX Spark N1X | 32GB unified memory | No compatible model listed; dedicated-VRAM setting set to 16GB | Qwen 3.6 35B-A3B at IQ4*XS ran through LM Studio after extensive adjustments; context-size crashes were reported |
| ASUS Zenbook A16 (2026) with Qualcomm Snapdragon X2 Elite Extreme | 48GB RAM | Qwen 3.8 27B, Nemotron 3.5 Bolt, and Qwen 3.6 35-A3B offered for one-click installation | The tested model ran on the CPU after Adreno driver errors |
The RTX Spark ran a model through LM Studio
The empty Hermes list did not prevent a separate local-model run on the OmniBook. Campanale got Qwen 3.6 35B-A3B running through LM Studio at IQ4*XS, a heavily compressed 4-bit format, after extensive adjustments. He also reported repeated crashes when changing the context size, making that route unreliable in his setup.
Speed and the coding task
Campanale said the RTX Spark was faster when a model fit. On the Zenbook, Adreno driver errors led the tested model to run on the CPU instead of the GPU; simple prompts took a few minutes and did not crash in the prompts he tried.
For a coding task, Campanale asked the Zenbook to write a Python script, run pytest, and revise the code until the tests passed. The task was still running after about an hour, when he stopped to move on to another test. On the OmniBook, the model crashed when asked to handle that task through Hermes.
Which tested setup fits your workflow?
- Choose the Zenbook configuration if one-click model access in Hermes is your priority. It offered three named models in Campanale’s test, though its tested model ran on the CPU and simple prompts took minutes.
- Consider the RTX Spark configuration if speed matters for models that fit and you are prepared to adjust the setup. Campanale reported faster performance when a model fit, but his 32GB OmniBook had no compatible model listed in Hermes and needed substantial tweaking to run Qwen through LM Studio.
Windows-on-Arm compatibility beyond local AI
In a separate Snapdragon X2 Elite Extreme analysis, Tom’s Hardware reported that AutoCAD was unsupported and that some games crashed or rendered incorrectly under Windows-on-Arm emulation. Those observations concern broader software compatibility, beyond Campanale’s local-AI workflow. Check support for the specific applications and games you rely on when evaluating a Snapdragon X2 laptop.
