Most AI asks you to send your most sensitive material to a company you have no relationship with, held on hardware you will never see, under a policy that can change next quarter. For a great many professionals that is not a trade-off they are permitted to make.
The alternative has always been obvious — keep the model on your own machine — and has always been impractical, because the models that were worth running would not fit. That is the problem the research exists to solve. Compression is not the point; it is what turns local-first from a preference into an option.
So the result needs somewhere to land. Float Point is that place: the same work, carried the rest of the way to something a practice can actually use on a Tuesday morning.