Importance matrix (imatrix)
Calibration data guiding which weights to preserve at higher precision during quantization.
An importance matrix is produced by running sample text through a model and recording which weights matter most to its outputs. The quantizer then allocates precision accordingly. Files quantized this way are usually marked with an IQ prefix or an i1 tag in the repository name.
The benefit is real but not systematically quantified: there is no published dataset of quality loss per quantization per model, so claims of precision here are guesses.