39 GB is very small, DeepSeek R1 without quantization at full context size needs almost a full TB of RAM/VRAM.

The large models are absolutely massive and you will still find some crazy homelabber that does it at home.