AI models store their knowledge as huge tables of numbers. Quantisation reduces how many digits each number uses, for example moving from high precision to a compact form, which shrinks the model and speeds up the maths with only a small loss in quality when done carefully.
The payoff is that capable models can run on phones, laptops and modest servers instead of only large machines, cutting cost and energy. The risk is that cutting precision too far degrades answers, so teams test the smaller version against the original and often protect the most sensitive parts of the model from aggressive reduction.