## TL;DR
Model compression reduces inference cost for deployment on resource-constrained devices. The three pillars — pruning, quantization, and distillation — can be combined for 10x+ compression with minimal accuracy loss.

## Core Explanation
Pruning removes weights (unstructured) or entire neurons/channels (structured) based on magnitude or gradient criteria. Quantization converts floating-point weights to lower precision (INT8, INT4). Knowledge distillation trains a small student model to mimic a large teacher.

## Detailed Analysis
Post-training quantization (PTQ): calibrate on representative data, no retraining. Quantization-aware training (QAT): simulate quantization during training for higher accuracy. GPTQ and AWQ enable 4-bit quantization of LLMs. DistilBERT achieves 97% of BERT's performance with 40% fewer parameters via distillation.

## Further Reading
- PyTorch: Quantization Tutorial
- Hugging Face Optimum
- Papers With Code: Model Compression