Q4_K_M Quantization Label Masks Varying Bit-per-Weight Across Model Versions

A technical observation has surfaced showing that models sharing the same Q4_K_M quantization label can differ significantly in actual bits per weight, ranging from 5.02 to 5.27. The Q4_K_M tag, commonly used in the machine learning community to describe a specific level of model compression, does not guarantee a uniform quantization level. This inconsistency means users selecting models based on the label alone may unknowingly load models with different memory and performance characteristics. The finding highlights a lack of standardization in how quantization labels are applied across model releases. The discussion, posted on Hacker News, has drawn early community attention to the need for more precise quantization metadata.