Strange that in the prior art they didn't list DwarfStar, as it is able to run the same model (probably quantized differently though) in less memory. Maybe the author isn't aware of it?