dealignai/Qwen3.8-Flash-Next-CRACK-JANG_4M
Abliterated vs JANGQ-AI/Qwen3.8-Flash-Next-JANG_4M. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 26 of 48 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: other 0/17, Embeddings 0/391, Attention 78/1066, MLP 21/1277, Norms 0/227, lm_head 0/3