I was tired of letting the OOM killer do my VRAM math, so I wrote a one-file CLI

My workflow for trying a new local model used to be: download 4 GB, watch the progress bar, then read...