tvall43/Qwen3.6-35B-A3B-heretic
Abliterated vs Qwen/Qwen3.6-35B-A3B. Looks like abliteration (refusal removal): only the output projections (o_proj, out_proj) changed, in 19 of 40 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/1, Attention 19/280, MLP 0/280, Norms 0/131, lm_head 0/1
12 derivatives checked
- AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4: Quantization
- Abiray/Qwen3.6-35B-A3B-heretic-GGUF: Quantization
- ENOSYS/Qwen3.6-35B-A3B-heretic-750-v1-GGUF: Quantization
- tvall43/Qwen3.6-35B-A3B-heretic-gguf: Same weights, converted
- TheCluster/Qwen3.6-35B-A3B-Heretic-MLX-mixed-9bit: Quantization
- tvall43/Qwen3.6-14B-A3B-FableVibes: LoRA-style
- CCSSNE/tvall43-Qwen3.6-35B-A3B-heretic-gguf: Same weights, converted
- TheCluster/Qwen3.6-35B-A3B-Heretic-MLX-mixed-3.9bit: Quantization
- TheCluster/Qwen3.6-35B-A3B-Heretic-MLX-4bit: Quantization
- tvall43/Qwen3.6-14B-A3B-VibeForged-v2: LoRA-style
- tvall43/Qwen3.6-14B-A3B-vibetuned: LoRA-style
- tvall43/Qwen3.6-14B-A3B-VibeForged: LoRA-style
3 other repos hold byte-for-byte identical weight files.