huihui-ai/Huihui-Qwen3-VL-4B-Instruct-abliterated
Abliterated vs Qwen/Qwen3-VL-4B-Instruct. Looks like abliteration (refusal removal): only the output projections (down_proj, o_proj) changed, in 36 of 36 layers. Abliteration edits exactly these matrices to remove a "refusal direction"; everything else is untouched.
Tensors changed: Embeddings 0/4, Norms 0/249, MLP 36/220, Attention 36/240
10 derivatives checked
- noctrex/Huihui-Qwen3-VL-4B-Instruct-abliterated-GGUF: Same weights, converted (fewer layers)
- mradermacher/Huihui-Qwen3-VL-4B-Instruct-abliterated-GGUF: Same weights, converted (fewer layers)
- mradermacher/Huihui-Qwen3-VL-4B-Instruct-abliterated-i1-GGUF: Quantization (fewer layers)
- noctrex/Huihui-Qwen3-VL-4B-Instruct-abliterated-i1-GGUF: Quantization (fewer layers)
- EZCon/Huihui-Qwen3-VL-4B-Instruct-abliterated-mlx: Same weights, repacked
- EZCon/Huihui-Qwen3-VL-4B-Instruct-abliterated-8bit-mlx: Quantization
- EZCon/Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-mlx: Quantization
- EZCon/Huihui-Qwen3-VL-4B-Instruct-abliterated-4bit-g32-mxfp4-mixed_4_8-mlx: Quantization
- yotisstudios/RaifuWars-Warrior-Qwen3VL-4B-v2: LoRA-style
- yotisstudios/RaifuWars-Warrior-Qwen3VL-4B: LoRA-style
6 other repos hold byte-for-byte identical weight files.