saidutta69/Qwen3-8B-heretic
Qwen3-8B-heretic by saidutta69 is an 8.2 billion parameter Qwen3 variant, decensored using the Heretic v1.4.0 abliteration method. This model suppresses refusal behavior via targeted weight edits, preserving the base model's knowledge and instruction-following capabilities. It is optimized for developers seeking Qwen3's dual-mode architecture without refusal guardrails, making it suitable for local agents, roleplay, and studying refusal mechanisms.
Loading preview...
Overview
saidutta69/Qwen3-8B-heretic is an 8.2 billion parameter large language model based on the Qwen3-8B architecture. Its primary differentiator is the suppression of refusal behavior through a technique called abliteration, specifically using Heretic v1.4.0. Unlike fine-tuning, abliteration directly edits specific weight directions responsible for refusal, leaving the base model's core knowledge and instruction-following largely intact. This results in an exceptionally low KL divergence of 0.0366 from the base model, indicating minimal degradation of original capabilities while significantly reducing refusals.
Key Capabilities
- Decensored Output: Suppresses refusal behavior, allowing the model to comply with requests that the base Qwen3-8B would typically refuse.
- Preserved Base Model Intelligence: Maintains the Qwen3-8B's knowledge and instruction-following, including its
<think>and direct-answer dual-mode architecture. - Efficient Modification: Achieves refusal suppression through targeted weight edits rather than extensive fine-tuning, which can often degrade coherence.
- Low KL Divergence: Exhibits a KL divergence of only 0.0366 from the base model, demonstrating that the modification is highly precise and minimally impacts other model functions.
Good For
- Local Agents: Ideal for developers building local AI agents where uninhibited responses are desired.
- Roleplay Scenarios: Suitable for applications requiring creative and unrestricted conversational or narrative generation.
- Studying Refusal Mechanisms: Provides a valuable tool for researchers to analyze and understand refusal behaviors in reasoning-capable models.
- Resource-Efficient Deployment: Available with various GGUF quantizations (Q4_K_M, Q5_K_M, Q6_K, Q8_0) to run on a range of consumer GPUs and CPU-only setups.