WaveCut/Qwythos-9B-v2-Heretic

WaveCut/Qwythos-9B-v2-Heretic: Decensored Language Model

This model is a 9 billion parameter, decensored version of empero-ai/Qwythos-9B-v2, created using the Heretic tool. Heretic is an automatic refusal-direction ablation tool that removes refusal behavior by surgically modifying the model's residual stream, specifically across attn.o_proj and mlp.down_proj layers. This process ensures that the model's core reasoning capabilities remain intact, with a reported KL divergence of 0.000712 against the base model, indicating a highly precise modification.

Key Capabilities & Features

• Decensored Output: Provides responses without the refusal behavior present in the original base model.
• High Fidelity: Ablation process maintains the original model's output distribution, ensuring minimal "damage" to its inherent capabilities.
• Qwen3.5 Hybrid Architecture: Utilizes a sophisticated architecture combining attention layers with linear/SSM (Mamba-style) layers.
• Extended Context Window: Supports a 32768-token context window, suitable for processing longer inputs.
• Optimized for Reasoning: The base model was post-trained on over 500 million tokens for deep chain-of-thought reasoning.

Good For

• Unrestricted Content Generation: Ideal for use cases where safety-alignment or refusal behaviors are undesirable.
• Research into Model Alignment: Useful for studying the effects of decensoring and understanding refusal mechanisms.
• Creative Applications: Suitable for generating diverse and unconstrained text outputs.

Quantized versions are available for various platforms, including GGUF (for llama.cpp, Ollama, LM Studio) and MLX (for Apple Silicon).