MoonRide/Qwen3.6-27B-heretic-ara
Abliterated vs Qwen/Qwen3.6-27B. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj, out_proj) changed, in 47 of 64 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, MLP 47/304, Attention 47/556, lm_head 0/1, Norms 0/319
1 derivatives checked
- MoonRide/Qwen3.6-27B-heretic-ara-GGUF: Same weights, converted