neopolita/Qwen3.6-23B-A3B-Niwaki-v2.3-4bit-mlx
Abliterated vs mlx-community/Qwen3.6-35B-A3B-8bit. Looks like abliteration (refusal removal): only the output projections (down_proj) changed, in 40 of 40 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/3, Norms 0/131, Attention 0/660, MLP 173/720, lm_head 0/3
1 derivatives checked
- neopolita/Qwen3.6-23B-A3B-Niwaki-v2.3-4bit-gguf: Quantization