Decoder. plain-English AI glossary

Quantization

● Core

Converting model weights and activations from high precision (float32) to lower precision (int8, int4) to save memory and speed up computation.

Think of it like

Storing a photo in lower resolution to take up less disk space.

Example

A model with 7 billion float32 weights uses 28GB of memory. Quantize to int8 (1 byte per weight) and it's 7GB.

How it actually works

Quantization trades accuracy for speed and memory. Most of the loss happens in the first few bits; going from float32 to float16 hurts less than float16 to int4. Post-training quantization is easy but less accurate; quantization-aware training is harder but better. Modern quantization methods (GPTQ, AWQ) are surprisingly good.

For product teams

Quantization multiplies your serving capacity. A 7B model that wouldn't fit on one GPU now fits.

For engineers

Start with int8 post-training quantization. If that's too lossy, try QAT or mixed-precision.

Key compression technique.

Read anything AI without the jargon

Look up any term in plain English, or save terms as you read with the free Chrome extension.

Open Decoder Add to Chrome