Quantization format

Q3_K_L

Files labelled Q3_K_L average 4.368 effective bits per weight across 2134 real quantizations — not the nominal 3.4375. That is 27% more than the label implies, because a quantization is a mixture: some tensors are always kept at higher precision.

From the file· 2134 files measured
Nominal bpw
3.4375
from the block layout
Measured average
4.368
2134 files
Range
1.647–23.759
varies by architecture
File sizes
0.02 GiB+
up to 454.05 GiB

Real files

one per model, most downloaded first
ModelSizeEffective bpwvs nominal
Qwen3.6-35B-A3B16.56 GiB3.956+15%
gemma-4-26B-A4B-it12.29 GiB3.978+16%
gemma-4-12B-it6.12 GiB4.392+28%
Qwen3.5-4B2.48 GiB4.576+33%
Hy3133.25 GiB3.831+11%
Qwen3-Coder-30B-A3B-Instruct13.58 GiB3.821+11%
gemma-4-31B-it15.66 GiB4.300+25%
Llama-3.2-1B-Instruct0.68 GiB4.742+38%
gemma-4-E2B-it3.06 GiB5.138+49%
Qwen3-8B4.13 GiB4.328+26%
Qwen3.5-0.8B0.49 GiB4.844+41%
Qwen3-VL-8B-Instruct-abliterated-v14.13 GiB4.044+18%
UI-TARS-1.5-7B3.81 GiB3.944+15%
Qwen3-4B2.09 GiB4.455+30%
Llama-3.1-8B-Instruct4.03 GiB4.306+25%
Ornith-1.0-35B15.73 GiB3.898+13%
Llama-3.2-3B-Instruct1.69 GiB4.520+32%
ThinkingCap-Qwen3.6-27B14.23 GiB4.468+30%
Wan2.1-T2V-1.3B1.15 GiB6.982+103%
Qwythos-9B-v24.89 GiB4.349+27%
Qwen3.5-122B-A10B57.37 GiB3.939+15%
Qwen3-Coder-Next35.60 GiB3.838+12%
Qwen2.5-32B-Instruct16.06 GiB4.211+23%
gemma-3-1b-it0.70 GiB6.013+75%
Qwen3.5-35B-A3B16.56 GiB3.956+15%
From the filewhat these mean

Other quantizations