Intel 258v - What's the point?(AI)

A user testing local AI model performance on the Samsung Galaxy Book 5 360 Pro with an Intel 258v processor reports disappointing results, achieving only about 30 tokens per second decode and 200 tokens per second prefill when running a Qwen 3.6 35B A3B model via llama.cpp with Vulkan, while noting that Intel's OpenVINO toolkit remains immature and underperforms compared to llama.cpp.

I recently lucked into a Samsung Galaxy Book 5 360 Pro talk about a mouthful with an Intel 258v 32GB model. I wanted to see how capable it was running local AI models and I’ve been kinda disappointed. Is there something I’m missing? Ubuntu Server 26.04 - Kernel 7.2 OpenVINO - So incredibly immature, I can run Qwen 3.6 35B A3B if I pull the latest nightlies and build everything myself but it just sucks in comparison to llama.cpp performance wise. llama.cpp - I tried Sycl and Vulkan, Vulkan was the clear winner with double the tokens per second output. I thought this generation was supposed to be more capable of running local models? Obviously there’s limitations with RAM amounts but parsing an A3B model and only getting roughly 30 t/s decode and 200 t/s prefill honestly this is the more painful part really sucks.