Qwen3.8-27B abliterated — NInfer artifacts - a Barding-Defense Collection

updated Sep 4

Refusal-removed Qwen3.8-27B (huihui-ai, layers 18-51) for the NInfer engine. NVFP4 and groupwise-int profiles. Format conversion only.

Barding-Defense/Qwen3.8-27B-huihui-abliterated-NVFP4-NInfer

Image-Text-to-Text • Updated Sep 4 • 21k • 11

Note Recommended. 21.5 GB, 19.0 GiB VRAM. 269.9 tok/s MTP decode, 88.8% acceptance. Needs an external llm-compressor stage to build.
Barding-Defense/Qwen3.8-27B-huihui-abliterated-groupwise-int-NInfer

Image-Text-to-Text • Updated Sep 4 • 31.5k • 1

Note Smaller: 18.2 GB, 15.9 GiB VRAM. 199.6 tok/s MTP decode, 86.2% acceptance. Builds in one command, no external quantizer and no patch.
Barding-Defense/Qwen3.8-27B-OBLITERATED-NVFP4-NInfer

Image-Text-to-Text • Updated Sep 4 • 853

Note Different source checkpoint (OBLITERATUS). 21.5 GB, 19.0 GiB VRAM. 248.7 tok/s MTP, 77.8% acceptance. Measured 0/20 refusal on AdvBench. NOTE: 3 frontend files replaced with canonical - see card.