If you're using a MoE model like DeepSeek V3 the full model is 671 GB but only 37 GB are active per token, so it's more like running a 37 GB model from the memory bandwidth perspective. If you do a quant of that it could e.g. be more like 18 GB.