← sources

Has anyone done a in-depth comparison between Qwen-3.8 27b @ 8bpw and Qwen-3.8-FN @ 4bpw?

source: reddit (reddit:LocalLLaMA)published 2026-11-03 15:02ingested 2026-11-04 02:41

retained text

Title says it all, I run dual 3090s (check out my post here for my [setup](https://old.reddit.com/r/LocalLLM/comments/1wtxrrv/sharing_my_qwen3827b_at_8bit_on_2x_rtx_3090_with/)) at int8A16 and have been wondering if I'm missing out with QFN. There's posts ad-nauseam on reddit with low bpw strata solutions that run on heterogeneous systems, however those don't interest me - I want high precision. I'm considering selling my setup and getting a DGX Spark, since that can run QFN at nvfp4 with a large context window. My use case isn't coding but generic agentic use case - I use hermes as a personal assistant and with openwebui. 

If anyone's able to run both and do a comparison, please let me know!

document fetched 2026-11-04 00:41 · canonical https://reddit.com/r/LocalLLaMA/comments/1xmv9ju/has_anyone_done_a_indepth_comparison_between