Skip to content
Technology

Quantization

A technique that reduces AI model size and speeds up inference by representing model weights with fewer bits (e.g., 4-bit instead of 32-bit). Enables running large models on smaller devices with minimal quality loss.

Why it matters

Quantization is why you can now run capable AI models on a laptop or even a phone instead of a data center. By storing each weight with fewer bits, it shrinks a model's memory footprint and speeds up responses, usually with only a small quality trade-off. This unlocks private, offline AI and dramatically lowers costs, which matters for anyone wary of sending data to the cloud.

A concrete example

A developer wants to run a large open-weight chatbot on a gaming PC with a single graphics card. The full model won't fit, so they download a 4-bit quantized version that does, and it answers nearly as well while using a fraction of the memory. A tip: if a quantized model feels noticeably less accurate, try a higher bit level like 6-bit or 8-bit as a compromise.

How to use it

It is what lets a large model run on hardware that could not otherwise hold it, at some cost in precision. If you are running models locally, this is the main lever between "fits and is fast" and "more accurate": try the smaller precision first and only move up if quality on your actual task is visibly worse. The loss is rarely uniform, so test on the work you care about rather than on a benchmark.

The common mistake

Assuming the quality cost is negligible because a benchmark score barely moved. Aggregate scores hide degradation on specific capabilities, and reasoning and long-form coherence tend to suffer before simple recall does.

Related terms

Put Quantization into practice

Access 750+ AI models and 60+ tools through Vincony — start free with 100 credits.