Back to glossary

Quantization

Nicemleme

D2 · Generative AI, capabilities and limitations

In model compression, quantization represents model values with fewer bits or a reduced set of numerical values.

Review status: 2026-12-05

Technical explanation

Weight quantization can reduce the memory needed to store parameters; GPTQ is one post-training method.

Conceptual boundaries

Fewer bits per weight is different from fewer parameters. This entry concerns model values, not every use of quantization in data processing.

Provider-neutral example

Illustrative: an engineer compares a lower-bit model with its original on the same task and device.

Limitations

Accuracy may decrease; compression alone does not establish latency or quality on a particular device.

Related concepts

Atomic claims and evidence

  1. 1.1Google Machine Learning Glossary: quantization
    Source
    Google Machine Learning Glossary: quantization
    Source role
    Authoritative source
    Exact locator
    quantization: reducing bits for parameters; possible decrease in prediction correctness
    Supported claim
    Google describes model quantization as reducing bits used for parameters and notes possible losses in prediction correctness.
    Last verification
    Review due
    Scope limitation
    This uses the parameter-compression meaning of an overloaded term.
  2. 2.1GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers
    Source
    GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers
    Source role
    Authoritative source
    Exact locator
    Abstract: weight quantization to 3 or 4 bits and evaluated compression results
    Supported claim
    GPTQ evaluates post-training weight quantization at low bit widths.
    Last verification
    Review due
    Scope limitation
    Its measured results do not apply automatically to every model or hardware configuration.