What is this compression technique called?

Quantization reduces the precision used to store each of the billions of numbers, called parameters, that make up an AI model's knowledge, for example converting numbers stored with 16 bits of precision down to just 4 or 8 bits. Because each number then takes less memory and computation, the model becomes smaller and faster to run, often making it possible to fit a capable model on ordinary hardware like a laptop or phone instead of specialized servers.

Buffering means temporarily storing data, often audio or video, for smooth playback, a different concept from reducing a model's numerical precision. Indexing means organizing data, like in a database, so it can be found quickly, unrelated to shrinking a model's number storage.

When done carefully, quantization can shrink a model's size dramatically while barely affecting its accuracy, which is why it has become a key technique enabling surprisingly powerful AI to run on everyday consumer devices rather than only in large data centers.