sasa2000/Qwen3-30B-A3B-Instruct-2507-REAM-heretic

Abliterated vs SamsungSAILMontreal/Qwen3-30B-A3B-Instruct-2507-REAM. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 46 of 48 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.

Tensors changed: Embeddings 0/1, Norms 0/193, MLP 798/2613, Attention 44/192, lm_head 0/1

3 derivatives checked

On the Hub · Downloads on Model Pulse