Quantization format

IQ3_S

IQ3_S has a nominal rate of bits per weight. We have no measured files carrying this label yet.

From the file· 235 files measured
Nominal bpw
from the block layout
Measured average
3.867
235 files
Range
2.884–20.569
varies by architecture
File sizes
0.03 GiB+
up to 377.51 GiB

Real files

one per model, most downloaded first
ModelSizeEffective bpwvs nominal
Qwen-Image-Edit-250922.43 GiB9.430
Qwen3-0.6B-Base0.36 GiB5.234
Mathstral-7B-v0.12.97 GiB3.517
Krea-2-Raw5.14 GiB3.441
Codestral-22B-v0.19.02 GiB3.484
Mistral-7B-v0.12.96 GiB3.516
Nanbeige4.2-3B1.87 GiB3.846
MN-12B-Mag-Mell-R15.18 GiB3.633
DeepSeek-Coder-V2-Lite-Instruct6.97 GiB3.814
Qwen3.6-12B-IQ-Ultra-Heretic-Uncensored-Thinking-V2-Hightop5.27 GiB3.729
Llama-3.2-3B-Instruct-uncensored1.59 GiB3.798
Gemma-4-31B-StyleTune12.22 GiB3.213
gemma-2-2b-it-abliterated1.27 GiB4.164
dolphin-2.9-llama3-8b3.43 GiB3.668
Hermes-4-14B6.23 GiB3.621
Qwen-Image-Edit8.36 GiB3.516
NEXUS-Medical0.71 GiB3.951
L3-8B-Stheno-v3.23.43 GiB3.668
Meta-Llama-3-8B3.43 GiB3.668
MN-Violet-Lotus-12B5.18 GiB3.633
MythoMax-L2-Kimiko-v2-13b5.45 GiB3.597
Hermes-3-Llama-3.1-8B3.43 GiB3.668
Qwen2-0.5B-Instruct0.32 GiB5.483
Mixtral-8x7B-Instruct-v0.119.03 GiB3.500
Qwen2.5-Coder-32B13.45 GiB3.525
From the filewhat these mean

Other quantizations