Single-User Inference Card

You could run Qwen 3.8 locally at almost 200 tok/s