Quantization

Storing model weights at lower numeric precision (e.g. 8-bit or 4-bit instead of 32-bit floats) to shrink memory use and speed up inference, at some cost to accuracy.