Model basics

What Is LLM Quantization?

Understand how lower-precision weights reduce memory use and introduce quality trade-offs.

Last reviewed 2026-09-22.

Quantization stores weights more compactly

Quantization changes how the model's numerical weights are represented. Instead of keeping every value at a higher precision, a quantized format uses fewer bits and a scheme for reconstructing useful values during inference.

Fewer bits usually means less weight memory. It does not mean the model has fewer parameters: it is a more compact representation of the same learned structure.

Bits per weight is a useful planning signal

Bits-per-weight gives a rough way to compare memory pressure. A 4-bit candidate generally needs less weight storage than an 8-bit candidate for the same model, although metadata, packing, and runtime buffers mean the final file is not a perfect parameter-count multiplication.

Names such as Q4_K_M, Q5_K_M, and Q8_0 identify particular quantization schemes. They should be treated as candidate formats, not as a universal quality ranking across every model family.

The practical trade-off

Higher-precision candidates usually preserve more numerical detail but require more memory. Lower-precision candidates can make a model fit on consumer hardware, but aggressive quantization can change outputs and may be a poor choice for a particular workload.

LLMGauge recommends among the candidates recorded for a model. Its recommendation is based on fit and the supplied bits-per-weight metadata; it is not a benchmark, blind quality test, or guarantee that one conversion is best for every task.

Continue with LLMGauge

References

These references support the concepts explained here. LLMGauge simplifies them for planning and does not present the guides as runtime documentation or performance benchmarks.