That's very large models at full quantization though. Stuff that will crawl even on a decent homelab, despite being largely MoE based and even quantization-aware, hence reducing the amount and size of active parameters.