It doesn't need extra quantization. The official weights are natively mixed precision FP4/FP8, so it fits in ~160GB. The API slowness is probably from being batched with other concurrent user requests. The provider's aggregate throughput gets higher but per-stream speed slows down.