All terms

Quantization

Storing and computing model weights with fewer bits (e.g., 8-bit or 4-bit) to shrink memory use and often speed inference.

Large Language Models1 min read

Definition

Quantization maps high-precision floats (FP16/FP32) to lower-bit integers or floats. Techniques include post-training quantization and quantization-aware training.

It enables running larger LLMs on smaller GPUs or CPUs with some quality tradeoff.

In simple terms

Storing music as a smaller compressed file — you save space and bandwidth, sometimes with a slight fidelity loss.

Where you see it

  • llama.cpp and GGUF quantized models on laptops.
  • INT8/FP8 inference servers in production.
  • QLoRA fine-tuning on a single consumer GPU.

How it works

  1. 1.Choose bit width

    8-bit, 4-bit, or mixed precision schemes.

  2. 2.Calibrate scales

    Map float ranges to integer grids.

  3. 3.Run inference

    Kernels compute with quantized weights.

Why it matters

  • Quantization is how open models become practical on real hardware budgets.

Often confused

  • Any quantization is lossless.

    Lower bits usually cost some quality — evaluate on your task.