# Importance Matrix in Low-Bit Quantization

An **importance matrix** (often called an "imatrix") is a technique used in post-training quantization—popularized by tools like **llama.cpp**—to improve the accuracy of aggressively quantized models (e.g., 2-bit, 3-bit, or mixed-precision formats like Q2_K, IQ4_XS).

## The Problem It Solves

Naive quantization treats all weights as equally important. But in reality:
- Some weights have a much larger impact on the model's output than others
- Quantization error on "important"