Quantization
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A compression technique that stores a model's weights in fewer bits (for example 32-bit floats down to 8-bit integers) to shrink size and speed up inference.
What is Quantization?
Quantization is a compression technique that stores a model’s weights, and sometimes its activations, in fewer bits – for example, converting 32-bit floating-point numbers (FP32) to 8-bit integers (INT8).
Each weight loses some numerical detail, but neural networks tolerate the rounding well. In exchange, the model takes a quarter of the memory and runs faster, since hardware moves and multiplies 8-bit integers far more cheaply than 32-bit floats.
Types of Quantization
- Post-training quantization converts a finished model’s weights to lower bit-widths without retraining. It is the fast path: minutes of work, at the cost of a small accuracy drop.
- Quantization-aware training simulates the low-bit arithmetic during model training, so the network learns weights that survive the conversion. It costs a training run but keeps accuracy closer to the original.
Either approach applies to weights alone or to weights and activations together. Quantizing both lets the entire forward pass run in integer arithmetic, which matters most on phones and edge devices without powerful GPUs.
Example of Quantization
A team builds an FP32 image classifier that weighs 400 MB and wants it running offline inside a mobile app. At that size it strains the app download and the phone’s memory.
Post-training quantization maps each FP32 weight to the nearest of 256 INT8 levels, scaled to the range of values in each layer. The file shrinks to roughly 100 MB – a 4x reduction, since each weight now needs 8 bits instead of 32.
On a benchmark set, accuracy slips from 91% to 90%. Inference on the phone speeds up several-fold, and the team ships the quantized model; the one lost point of accuracy bought a model that fits the device.
Related AI terms: Inference · Sparsity · Model Training · Edge Computing · Neural Network
Did you like the Quantization gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About