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 :)

It's better for VRAM poor people. I get 4-5 t/s with 27B and 20-30 t/s with 35B A3B.

anecdotes: 35B-A3B does want more memory, bigger model. But if you get it running it will be faster and more enjoyable to use -- text will fly by -- due to only 3B params being active, in my experience at least.

Runs faster on bad hardware.