Laguna S 2.1 Q4_K_M grows from 68 GB to 96 GB after keeping 8 layers FP16

BawbbySmith · reddit · 2026-07-29

A Reddit user noticed that Laguna S 2.1’s Q4KM GGUF file jumped from 68 GB to 96 GB.

They suspect the new build keeps 8 layers in FP16 while the rest remains 4-bit.
The post asks why the quantization strategy changed and whether it is meant to fix quality issues at that compression level.
The author also mentions looping issues at higher context lengths with similar quantized builds, including Unsloth’s version.

More from Infra