User benchmarks for the Intel Arc Pro B70 (32 GB VRAM) demonstrate local LLM inference performance using llama.cpp with the Vulkan backend. The tests specifically evaluate the Qwen3.6-27B (Q4_K_M) and Qwen3.6-35B-A3B models on a system powered by an Intel Ultra 7-265 CPU and 96 GB of DDR5 RAM.