Quantisation
Also: Quantization
Storing model weights with lower precision so that models become smaller and faster.
Instead of storing each weight with 16 or 32 bits, only 8 or 4 bits are used, for example. The model then needs only a fraction of the memory and runs on less powerful hardware.
The price is usually a small loss in quality. Quantisation is why useful language models can run on ordinary laptops today.
General information, not legal or professional advice.