OPTIMIZATION

Full Precision vs Quantized LLMs

Quantization compresses LLM weights to use less memory and run faster — but how much quality do you lose?

Full Precision vs Quantized

The same model, two different representations

FULL PRECISION

Standard Weights

Each weight stored as a full 32-bit floating-point number, preserving maximum precision and accuracy.

Data Type
FP32
Bits per Weight
32
7B Model Size
~28 GB
GPU Memory
~32 GB
QUANTIZED

Compressed Weights

Weights are mapped to fewer bits, dramatically reducing memory usage while maintaining acceptable quality.

Data Type
INT4 / INT8
Bits per Weight
4 – 8
7B Model Size
~4 – 8 GB
GPU Memory
~6 – 10 GB

Weight Representation: Same Model, Different Sizes

FP32 (Full Precision)
32 bits per value
INT8 (8-bit Quantized)
8 bits per value (4× smaller)
INT4 (4-bit Quantized)
4 bits per value (8× smaller)

Model Size by Quantization Level

Based on a 7-billion parameter model

FP32 (Full Precision) 28 GB
100%
FP16 (Half Precision) 14 GB
50%
INT8 (8-bit) 7 GB
25%
INT4 (4-bit) 3.5 GB
12.5%

Quantization Methods

Different approaches to compressing model weights

FP16 / BF16

Half-precision or brain-float 16. Simple reduction from FP32 with minimal quality loss. Widely supported.

Near Lossless

INT8 Quantization

Maps weights to 256 discrete levels. 4× compression from FP32 with minimal quality degradation.

Minimal Loss

INT4 Quantization

Maps weights to 16 discrete levels. 8× compression. Some quality loss but still highly usable.

Moderate Loss

GPTQ / AWQ

Advanced methods that calibrate quantization per-layer using a small dataset. Better quality at same bitwidth.

Smart Loss

Quality vs Compression Trade-off

FP32
FP16
INT8
GPTQ-4
INT4
← Maximum Quality Maximum Compression →

Next in Series

KV Cache

Speed, memory trade-offs, and inference optimization