Quantization
Shrinking a model by storing its numbers with less precision, trading a little accuracy for much lower cost.
A model’s parameters are usually stored as high-precision decimals; quantization rounds them to coarser values, so the same model takes up a quarter of the memory or less. Done carefully, the loss in answer quality is small, which is why quantized models made it practical to run capable systems on ordinary computers. The technique sits behind many of the announcements about models that run on a laptop.