Newer models such as Phi2 run comfortably with 4GB and are good enough to be useful for casual interaction. Sticking with local inference, multiple small models

Newer models such as Phi2 run comfortably with 4GB and are good enough to be useful for casual interaction. Sticking with local inference, multiple small models tuned for specific usage scenarios is where it's at.