Quantization

Quantization lowers the precision used to represent a model's numbers. Instead of storing weights and activations as 32-bit floating point, it maps them to a smaller format such as 8-bit integers. This shrinks the model, reduces memory bandwidth, and speeds up computation on hardware that runs low-precision math efficiently, usually with only a small drop in accuracy.

It is one of the most widely used techniques for putting large models into production and onto edge devices.

Values are mapped from a high-precision range to a smaller set of levels, with a scale that preserves as much information as possible. Post-training quantization applies this to an already-trained model, while quantization-aware training simulates the lower precision during training so the model adapts and loses less accuracy. The best approach depends on how aggressive the precision reduction is.

Real deployments live under memory, latency, and power constraints, and quantization is often what makes a strong model actually shippable. It pairs naturally with knowledge distillation, and together they turn large research models into efficient systems, which is critical for on-device visual and physical AI.

It can, but usually only slightly, especially with 8-bit precision or quantization-aware training. More aggressive reductions trade more accuracy for more efficiency.

It simulates low-precision arithmetic during training so the model learns to tolerate it, typically preserving more accuracy than quantizing after the fact.

Knowledge distillation, Backbone, Convolutional Neural Network (CNN), Fine-tuning.