Comment by pickledish

The "a3b" refers to its active parameters -- unlike 27b it is a mixture of experts model, so it runs much faster, about as fast as a 3b model, but needs as much memory as a 35b model! So good for unified memory systems like macs :)