Qwen3.6-27B - a huginnfork Collection

huginnfork 's Collections

updated 9 days ago

FP8 & NVFP4 quants of Qwen/Qwen3.6-27B, plus the heretic-v2 uncensored line. Vision tower, SSM block and MTP head kept in bf16.

huginnfork/Qwen3.6-27B-NVFP4A16

Image-Text-to-Text • 20B • Updated Apr 27 • 220 • 1

Note Plain base, NVFP4 W4A16, ~28.5 GiB. KLD 0.0747 nats, PPL 7.652 (+4.74 %) vs bf16 base.
huginnfork/Qwen3.6-27B-uncensored-heretic-v2-mtp

Image-Text-to-Text • 28B • Updated Apr 27 • 158 • 4

Note bf16 source for the uncensored line — heretic+ARA abliteration with the MTP head regrafted from upstream. KLD 0.0425 nats, PPL 7.462 (+2.13 %) vs base. ~51 GiB.
huginnfork/Qwen3.6-27B-uncensored-heretic-v2-mtp-FP8

Image-Text-to-Text • 28B • Updated 19 days ago • 1.99k • 5

Note FP8 attnbf16 — MLPs only, the whole self_attn path stays bf16. KLD 0.0203 nats vs the bf16 parent, 73.0 % vLLM MTP draft acceptance. Best quality of the uncensored line.
huginnfork/Qwen3.6-27B-uncensored-heretic-v2-mtp-NVFP4A16

Image-Text-to-Text • 20B • Updated Apr 27 • 148 • 2

Note NVFP4 W4A16 — smallest uncensored build, ~28.5 GiB. KLD 0.1012 nats, PPL 7.802 (+6.79 %) vs base.