saidutta69/Qwen3-8B-heretic

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.