Benchmarks for the Intel Arc Pro B70 (32GB Battlemage) using llama.cpp with a Vulkan backend show stable performance running Qwen3.6-35B-A3B. The hardware achieved approximately 130 t/s with 4-bit quantization fully in VRAM and 69 t/s with 8-bit hybrid offload. Testing was conducted on a Ryzen 9 9900X system using Intel driver 32.0.101.8860.
Read original
reddit/r/LocalLLM