llmfan46/Qwen3-VL-32B-Instruct-uncensored-heretic
Abliterated vs Qwen/Qwen3-VL-32B-Instruct. Looks like abliteration (refusal removal): only the output projections (o_proj) changed, in 4 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 0/316, Attention 4/364, lm_head 0/1
3 derivatives checked
- mradermacher/Qwen3-VL-32B-Instruct-uncensored-heretic-GGUF: Quantization (fewer layers)
- mradermacher/Qwen3-VL-32B-Instruct-uncensored-heretic-i1-GGUF: Quantization (fewer layers)
- llmfan46/Qwen3-VL-32B-Instruct-uncensored-heretic-GGUF: Same weights, converted (fewer layers)