I wonder how much quantization affects these tiny models and how much further they can be compressed