Quantization reduces the number of bits used per weight. Sixteen-bit weights take about two gigabytes per billion parameters; eight-bit takes roughly one; four-bit takes roughly half a gigabyte. For anyone trying to run a model on a card they already own, quantization is the difference between fitting and not fitting.
The cost is quality, and the cost is not uniform. A model at four bits may produce nearly identical output to its sixteen-bit parent on ordinary text and degrade noticeably on long chains of reasoning or on precise formatting. Two quantizations of the same model from different toolchains are not interchangeable, because they choose which weights to approximate differently.
A practical rule: quantize to fit your card, then re-run the one prompt you graded against the full-precision model before you trust the smaller build. If the answers diverge on the cases that matter to you, the memory you saved is not worth it.
Record the quantization and the file hash next to the download. A later file with a different hash is a different model, even when the card title has not changed. See open source and proprietary models.