QuIP Sharp

2-bit quantization method delivering near-fp16 quality — extreme model compression.