MY ONE BLOGby Manishdai
Artificial Intelligence

What is Quantization?

Quick answer

Reducing the numerical precision of an AI model's weights, for example from 16-bit to 4-bit numbers, to make it smaller and faster.

Quantization, explained

Neural network weights are usually stored as floating-point numbers with high precision. Quantization maps them to lower-precision formats, such as 8-bit or 4-bit integers. This can shrink a model's memory footprint several times over and speed up inference, often with only a small drop in accuracy.

Quantization is a key reason large language models can now run on laptops, phones, and smaller GPUs. It can be applied after training, called post-training quantization, or built into training, called quantization-aware training, which usually preserves accuracy better. Lower precision can affect some tasks more than others, so teams evaluate quantized models carefully. In signal processing, quantization more generally refers to converting continuous values into discrete levels.

Related terms

Inference

Using a trained machine learning model to make predictions or generate outputs on new data.

Knowledge Distillation

A training technique in which a smaller student model learns to imitate a larger teacher model, keeping much of its skill at lower cost.

Large Language Model

An AI model trained on vast amounts of text to understand and generate human-like language, code, and other content.

Neural Processing Unit

A specialized processor designed to run AI and machine learning tasks efficiently, often built into phones and laptops.

Guides that mention Quantization