Vikhrmodels/QVikhr-3-4B-Instruction
Abliterated vs Qwen/Qwen3-4B. Looks like abliteration (refusal removal): only the output projections (down_proj, embed_tokens, 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 1/1, Norms 0/145, MLP 36/108, Attention 36/144
4 derivatives checked
- Vikhrmodels/QVikhr-3-4B-Instruction-GGUF: Same weights, converted
- mradermacher/QVikhr-3-4B-Instruction-i1-GGUF: Quantization
- mradermacher/QVikhr-3-4B-Instruction-GGUF: Same weights, converted
- prithivMLmods/QVikhr-3-4B-it-F32-GGUF: Same weights, converted