Skip to content
AI Stack Nest
Concepts3 min read · updated 18 Nov 2026

What Quantization Means

A plain-language explanation of what quantization is, why it matters for running AI models, and what you trade away

If you have looked at AI models available for local use, you have probably seen files described with terms like "8-bit" or "4-bit." These labels describe something called quantization. It is one of those concepts that sounds technical but is based on a straightforward idea.

The Core Idea

Every value inside a trained model, the numbers that determine how it responds to any input, is stored as a number. By default, those numbers are stored with high precision, meaning each one takes up a fixed amount of space in memory. A model with many billions of such values takes up a large amount of storage and requires a lot of memory to run.

Quantization is the process of reducing the precision of those numbers. Instead of storing each value with full precision, you store a rounded, lower-precision version. The file gets smaller. The model loads faster and needs less memory to run. The trade-off is that rounding introduces a small amount of error into the values, which can slightly reduce the quality of the output.

Why This Matters Practically

A full-precision model might require more memory than most consumer computers have available. A quantized version of the same model might run on a laptop with a standard amount of RAM. This is the reason quantization exists as a practical technique: it makes models accessible to people who do not have specialised or expensive hardware.

The relationship is roughly: lower precision means smaller file, less memory required, faster generation, and some reduction in output quality. Higher precision means larger file, more memory required, slower generation, and better output quality.

Different Levels of Precision

You will see different levels of quantization described by the number of bits used to store each value. The original, full-precision format uses a relatively large number of bits per value. Quantized formats use fewer bits. The more you reduce the bits, the smaller and faster the model becomes, but the more information is lost in the rounding.

The difference in quality between a lightly quantized model and a heavily quantized one is often noticeable on tasks that require precise reasoning, nuanced writing, or handling complex instructions. For simple, everyday tasks, even a heavily quantized model may produce results that are practically good enough.

What You Are Actually Trading Away

It helps to think of it like this: imagine you have a detailed painting and you need to describe it in a limited number of words. With many words, you can be precise. With fewer words, you have to round off the details. The subject is still recognisable, but fine details are lost. Quantization does the same thing to the internal values of a model.

For most practical use, the trade-off is worth it. A quantized model that runs on your hardware is more useful than a full-precision model that does not run at all.

Worked Example

Suppose a model is available in two versions. The larger version runs only on hardware with a substantial amount of memory. The smaller, quantized version runs on a standard laptop with moderate RAM and generates text at an acceptable speed. For everyday writing tasks, note-taking, or simple questions, the output of the quantized version is barely distinguishable from the full-precision one. For tasks that need careful multi-step reasoning, the full-precision model may produce noticeably better results.

The practical advice: start with a quantized version to check whether your hardware can run the model at a useful speed. If the output quality feels insufficient for your use case, try a less aggressively quantized version if one is available.

More setup guides

Not sure your computer can run a model? Use the will-it-run checker.