Quantization format

UD_IQ3_XXS

UD_IQ3_XXS has a nominal rate of bits per weight. We have no measured files carrying this label yet.

From the file· 213 files measured
Nominal bpw
from the block layout
Measured average
3.298
213 files
Range
2.075–5.506
varies by architecture
File sizes
0.11 GiB+
up to 392.74 GiB

Real files

one per model, most downloaded first
ModelSizeEffective bpwvs nominal
Qwen3.6-27B11.17 GiB3.454
Qwen3.6-35B-A3B12.30 GiB2.940
Qwen3.5-9B3.74 GiB3.329
gemma-4-26B-A4B-it10.63 GiB3.441
gemma-4-12B-it4.32 GiB3.104
DeepSeek-V4-Flash95.93 GiB2.832
Qwen3.5-4B1.82 GiB3.346
gemma-4-E4B-it3.46 GiB3.719
Qwen3-Coder-30B-A3B-Instruct11.97 GiB3.367
gemma-4-31B-it11.02 GiB3.028
Qwen3-VL-30B-A3B-Instruct12.00 GiB3.318
Qwen-AgentWorld-35B-A3B12.80 GiB3.172
Llama-3.2-1B-Instruct0.54 GiB3.725
gemma-4-E2B-it2.21 GiB3.705
Qwen3-8B3.18 GiB3.331
Qwen3.5-0.8B0.37 GiB3.647
Qwen3-4B1.56 GiB3.330
GLM-5.2262.34 GiB2.991
Llama-3.1-8B-Instruct3.09 GiB3.309
Ornith-1.0-35B12.80 GiB3.172
Llama-3.2-3B-Instruct1.28 GiB3.412
Kimi-K2.7-Code351.00 GiB2.848
Qwen3.5-122B-A10B41.67 GiB2.862
Qwen3-Coder-Next26.53 GiB2.860
gemma-3-1b-it0.55 GiB4.734
From the filewhat these mean

Other quantizations