⚠ Switch to EXCALIDRAW VIEW in the MORE OPTIONS menu of this document. ⚠ You can decompress Drawing data with the command palette: ‘Decompress current Excalidraw file’. For more info check in plugin settings under ‘Saving’

Excalidraw Data

Text Elements

NF4 vs Standard Integer Quantization

Neural Network Weights Distribution

(Bell Curve / Gaussian)

-8

0

+7

Weight Values

Frequency

Most weights cluster near zero!

Standard Integer Quantization

-8

-6

-4

-2

0

+2

+4

+7

Equal-sized buckets (Uniform)

Wastes precision at extremes!

NF4 (NormalFloat 4-bit)

-8

0

+7

More buckets near zero (Bell-curve aligned)

Matches weight distribution!

Result:

NF4 captures ~99% of 16-bit precision

Standard 4-bit loses significant info

• Standard: Assumes uniform distribution

• NF4: Designed for Gaussian distribution

• More resolution where data lives = less loss