What is Quantisation?

Storing model weights at lower numeric precision to cut memory use and speed up inference, usually with a small quality cost.