coder3101/Qwen3-VL-32B-Thinking-heretic
Abliterated vs Qwen/Qwen3-VL-32B-Thinking. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 51 of 64 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/373, MLP 29/316, Attention 51/364, lm_head 0/1
1 derivatives checked
- SerialKicked/Qwen3-VL-32B-Thinking-heretic-GGUF: Quantization (fewer layers)