The goal is not (yet?) to be able to run those models on most of consumers devices (mobile, old laptops etc.), but at least to self-host the model on high-end c

The goal is not (yet?) to be able to run those models on most of consumers devices (mobile, old laptops etc.), but at least to self-host the model on high-end consumer GPU which is not possible right now. For now you need multiple specialized GPUs like nvidia V100/A100 with a high amount of VRAM, having such models to run on a single rtx40*/rtx30* would already be an achievement