Lfm2 feels like a pretty solid local models for how efficient it is. Still testing tho.