LLM infrastructure
LLM quantization calculator
Compare a higher-precision source model with a lower-bit target. Add a metadata allowance so the result is closer to a real quantized file than a nominal bit-width calculation.
Free · no sign-in · updates instantly
1. Enter the model assumptions
For example, 7 means 7 billion parameters
Baseline weight representation
Nominal quantized weight width
Scales, zero points, padding, and exceptions
Results are weights-only estimates.
Source weights0 GB
Before quantization
Effective target0 bits/param
Target + metadata
Target weights0 GB
After quantization
Memory saved0 GB · 0%
Compared with source
Limit: add KV cache, activations, runtime buffers, and other loaded components when sizing a GPU. This estimate uses decimal GB.
How the estimate works
Raw weights use parameters × bits ÷ 8. The target bit width is multiplied by the metadata allowance before the target memory is calculated. For a 7B model, 16-bit source weights use 7 × 16 ÷ 8 = 14 GB; a 4-bit target with 12.5% overhead uses 4.5 effective bits and about 3.94 GB.
effective bits = target bits × (1 + metadata % ÷ 100)
weight memory (GB) = parameters (billions) × effective bits ÷ 8
Read the full explanation
See quantization explained for quality, speed, cache, and runtime trade-offs. For complete deployment sizing, read how to estimate GPU VRAM.