User tests Qwen3, Qwen3.6, and Gemma4 on Intel 285HX CPU

A user benchmarks Llama.cpp with Qwen3, Qwen3.6, and Gemma4 models on an MS-02 mini-pc equipped with an Intel Core Ultra 285HX processor and 64GB of RAM, noting the absence of a dedicated GPU.

• CPU-only inference yielded the highest speeds, reaching up to 16.5 tokens per second for Qwen3-30B-A3B-Instruct.
• SYCL backend performance ranged from 8 to 12 tokens per second across the tested models.
• Vulkan backend using the integrated GPU was significantly slower, producing only 0.5 to 4 tokens per second.

The user concludes that CPU-only execution is currently the fastest option for this specific hardware configuration and questions why Vulkan underperformed compared to SYCL.