lflare/Qwen3.8-Flash-Next-NVFP4-Abliterated
Text Generation • 94B • Updated • 1.46k • 8
Qwen3.8-Flash-Next Abliterated
In-place refusal-direction edits of Qwen3.8-Flash-Next NVFP4. No tensor changes format — the edit is applied on the quantisation grid itself.
Note The NVFP4 checkpoint, edited in place. main = tag v3: biproject at decoder layer 43, thinking off, 25,188 of 25,188 tensors edited, 0 structural diffs. lavd 30/30 and estonia 30/30 at n=30. Tags v1 and v2 keep the earlier builds.
lflare/Qwen3.8-Flash-Next-NVFP4-MXFP8-CSF-QAD-Abliterated
Image-Text-to-Text • 97B • Updated • 460 • 2Note The NVFP4 + MXFP8 CSF container. main = tag v2, a rank-2 edit that removes the thinking-off and thinking-on refusal directions together at layer 42. 0/136 held-out harmful in both regimes, over-refusal 2/59 to 0/59, lavd 30/30, estonia 30/30. Needs a runtime with a CSF path — karmic-kraken-20261016.2 or newer.
lflare/Qwen3.8-Flash-Next-FLARE-Abliterated
Image-Text-to-Text • Updated • 197 • 3Note The Flare package build of the same rank-2 edit. The package is on branch gb202 and main carries the card only, so pass revision=gb202 to download weights. Tag v3 is the ABI-4 re-export.