What is Model Compression

Definition

Model compression is the set of techniques, including quantization, pruning, and distillation, that reduce a model's size and computational demands while preserving as much accuracy as possible, making models cheaper, faster, and deployable on constrained hardware.

« Back to Glossary Index

Reduces model size and compute for cheaper, faster inference
Enables deployment on edge and resource-limited devices
Lowers latency and energy use in production
Retains most accuracy through careful compression methods

Real World Example

A team compresses a large vision model with pruning and quantization so it runs in real time on a factory camera's embedded chip, achieving fast on-site defect detection that the full model could not deliver there.

FAQs

What techniques does model compression include?

Pruning removes redundant weights, quantization lowers precision, and distillation trains a smaller model to mimic a larger one.

Why compress models?

To cut cost, latency, and memory so models can run efficiently in production and on constrained devices.

Does compression reduce accuracy?

Some accuracy may be lost, but careful methods keep degradation small relative to the efficiency gained.