Glm 5, which i think is near frontier model capabilities, takes around 1.5 Tb of vram un-quantized.

DeepSeek’s latest version is 508GB, and even that is wasteful, compared to the wild difference in sizes with some of the quantized versions. The huge amount of parameters and weight sizes are completely unoptimized in terms of sizes, and there’s all kinds of diminishing returns with the haphazardous way these LLM corps are applying new model versions.