>In a year or so, the open source models will become good enough (in both quality and speed) to run locally.

Based on what?

And where? On systems < 48GB?