Generally about ~1gb ram per billion parameters. I've run a 30b model (vicuna) on my 32gb laptop (but it was slow).