4-bit, 8-bit, and FP16 memory are not the same kind of number

Theoretical bits/parameter, real GGUF files, and BF16 safetensors answer different questions.

Updated 2026-09-19. Figures use calculator 0.1.0 unless the article says otherwise.

The identity parameters × bits / 8 is a storage sketch. For 8,000,000,000 parameters at 4 bits it is exactly 4,000,000,000 bytes. That test exists in this repo. It is not a promise that every “4-bit” file is 0.5 bytes per parameter.

Qwen2.5-7B-Instruct (7.61B parameters on the model card):

Kind Bytes What it is
BF16 safetensors total_size 15,231,233,024 Selected BF16 tensors
Official GGUF Q8_0 shards 8,098,525,408 Those three Q8_0 files only
Official GGUF Q4_K_M shards 4,683,073,632 Those two Q4_K_M files only
Theoretical INT4 7.61e9 × 4 / 8 ≈ 3.805e9 Sketch, no metadata

Q4_K_M is not 4.000 bits. Block scales and non-quantized tensors make the file larger than the theoretical INT4 line. AWQ/GPTQ names are not used as 0.5 bytes/parameter without file metadata.

Calculator 0.1.0, dataset 2026-09-19.1. Context and KV are extra in every row.

Open the calculator