Comment by millimeterman

I suspect the community will start creating lower precision/quantized versions of the model very quickly. LLaMa 30b quantized to 4 bits is runnable on a 3090/4090.