TensorFlow supports weight quantization, but XNOR seems to be rewriting the trained model graph into a one that only uses binary boolean ops, and probably optimizing the graph along the way.