Quantization
Storing weights at reduced precision to shrink a model, trading some quality for memory.
A quantization label such as Q4_K_M names a mixture, not a uniform precision. Normalization and bias tensors stay at full precision, and attention and output tensors are commonly promoted to a higher rate to protect quality. The consequence is that the effective bits per weight of a real file is always above the nominal rate of its label — Q4_K_M averages closer to 5.4 than to 4.5 in practice.
This is why estimating a file's size by multiplying parameters by the nominal rate is wrong before you have started, and why we sum published file bytes instead.