Quantization
Quantization reduces the numerical precision of a model's weights and activations — for example from 16-bit floating point to 8-bit or 4-bit integers — to shrink memory use and speed up inference. It trades a small, often negligible, loss in accuracy for substantially lower cost and hardware requirements.
Definition
QuantizationNeural network weights are normally stored in 16- or 32-bit floating point. Quantization maps these values to a lower-precision representation, cutting the memory footprint and the bandwidth needed to move weights, which lets larger models run on smaller or cheaper hardware and often speeds generation.
Methods range from simple post-training quantization to more careful schemes that preserve sensitive weights or calibrate on sample data. Techniques such as GPTQ enable accurate low-bit quantization of large models. Aggressive quantization can degrade quality, so precision is chosen to balance size, speed, and acceptable accuracy loss.
Frequently asked questions
What is Quantization?
Quantization reduces the numerical precision of a model's weights and activations — for example from 16-bit floating point to 8-bit or 4-bit integers — to shrink memory use and speed up inference. It trades a small, often negligible, loss in accuracy for substantially lower cost and hardware requirements.
Sources
Related terms
← Full AI & prompt engineering glossary
Put Quantization to work with Prompeteer, the Agentic Contextual AI Platform →