[binary132]

I find the Q8 runs a bit more than twice as fast as gpt-120b since I don’t have to offload as many MoE layers, but is just about as capable if not better.