Intel Arc Pro B70 32GB Finally Works with Qwen3.5-27B—Here's the Real Speed Numbers

So you've got an Intel Arc Pro B70 GPU sitting on your desk and wondering if it's actually worth using for AI inference? This person spent weeks wrestling with vllm (a GPU acceleration framework) to get it working, and now they've got actual performance numbers to share. The key findings: both llama.cpp and vllm hit around 12 tokens per second on Qwen3.5-27B in Q4 quantization—which is decent but not earth-shattering. Interestingly, some optimization tricks like tensor parallel actually made things worse (probably due to their PCIe setup), while pipeline parallel helped with batch processing but hurt single queries. If you're thinking about Arc Pro for local AI inference, this is the real-world data you need.