Quantization format

MXFP4

MXFP4 has a nominal rate of bits per weight. We have no measured files carrying this label yet.

From the file· 13 files measured
Nominal bpw
from the block layout
Measured average
4.761
13 files
Range
4.113–8.501
varies by architecture
File sizes
2.84 GiB+
up to 478.42 GiB

Real files

one per model, most downloaded first
ModelSizeEffective bpwvs nominal
DeepSeek-V4-Flash145.29 GiB4.290
gpt-oss-20b11.28 GiB4.503
Krea-2-Turbo7.34 GiB4.920
gpt-oss-120b59.03 GiB4.211
Inkling478.42 GiB4.315
Huihui-Qwen3-VL-4B-Instruct-abliterated2.84 GiB5.496
Apertus-70B-Instruct-250969.87 GiB8.501
gpt-oss-safeguard-20b11.28 GiB4.503
gpt-oss-safeguard-120b59.03 GiB4.211
DeepSeek-V4-Flash-0731145.64 GiB4.113
From the filewhat these mean

Other quantizations