Quantization

The count stays the same — only the digits shrink

A model carries billions of numbers set during training. Normally these are stored with generous room down to several decimal places — enough to capture the finest differences, but each one eating up a lot of space.

Quantization moves these numbers onto a coarser scale. Anything that falls between two marks gets rounded to the nearest one. No number disappears, and no connection gets severed. The table keeps exactly as many cells as before — only the numbers written inside those cells get shorter.

Cut the space in half and total size drops by roughly half too. Cutting it to a quarter is common as well. A model that was over ten gigabytes coming down to four or five gigabytes is often the difference between "won't run" and "runs fine" on a personal computer.

Where you set the marks is what separates good from bad

Making the scale coarser means deciding what range of values to spread the marks across. If most numbers cluster near zero but the marks get stretched out evenly across a huge range, that crowded region near zero ends up crammed into a single mark, wiping out exactly the detail that mattered.

So the mark spacing usually gets set separately, layer by layer or even row by row. Sometimes a handful of unusually large outlier values get left at full precision while everything else gets coarsened. Even at the same digit count, how carefully this step is done can leave a real gap in output quality.

Often a small batch of sample data gets run through first just to measure where values cluster, and the marks get set from that. No retraining happens, so the whole process can take minutes rather than hours.

Why AI runs on laptops and phones

When running a model, the biggest time cost usually isn't the computation itself — it's pulling the numbers out of memory. Shorter numbers mean more of them can be moved in the same amount of time, so the speed of getting an answer noticeably picks up.

Whether everything fits in graphics memory also hinges on this. If a model can't fit, pieces have to keep swapping in and out, which drags speed down badly. Chat AIs that run right in your browser, translators that work with no internet connection, and photo-sorting features on your phone are, more often than not, running a model with a coarsened scale.

Knowing what gets given up

Getting lighter comes at a cost. Rounding pulls each number a little away from its original value, and that drift accumulates as it passes through layer after layer. Producing even a single line of an answer passes through dozens of layers, so a gap that's invisible at the start becomes noticeable by the end.

Judgment calls that were close to begin with feel it the most. It shows up clearly in things like long arithmetic or code, where a single wrong character throws off the whole result, and in answers that run very long. In casual conversation or summarizing, on the other hand, the difference is often hard to notice at all.

A different move from pruning

Since both make a model lighter, quantization and pruning often come up together — but they touch different things. Quantization never throws away a single number; it leaves every one in place and only coarsens how they're written. Pruning does the opposite — it removes unused connections outright, cutting down the count.

Neither gets in the other's way, so they're often combined: trim away unused connections first to shrink the count, then coarsen the scale of what's left.