saidutta69/Qwen3-8B-heretic

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 16, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

Qwen3-8B-heretic by saidutta69 is an 8 billion parameter, 32k context length Qwen3 variant that has been decensored using directional ablation (abliteration) to suppress refusal behavior. This model retains the base Qwen3's knowledge and instruction-following capabilities, including its thinking/non-thinking dual-mode architecture, but with significantly reduced refusals. It is optimized for developers seeking a less restrictive model for local agents, roleplay, and studying refusal mechanisms.

Loading preview...

Qwen3-8B-heretic: Decensored Qwen3 Variant

This model is a decensored version of the 8 billion parameter Qwen3-8B, created by saidutta69 using the Heretic v1.4.0 tool. Unlike traditional fine-tuning, this variant achieves refusal suppression through "abliteration," a method that involves targeted weight edits to the attention output and MLP down-projections. This approach minimizes impact on the base model's core knowledge and instruction-following abilities, preserving its original coherence and dual-mode architecture (supporting both <think> and direct-answer modes).

Key Differentiators

  • Decensored Behavior: Refusal mechanisms are significantly suppressed, with refusals dropping from 100/100 to 10/100 on adversarial prompts, while maintaining an exceptionally low KL divergence of 0.0366 from the base model.
  • Abliteration Method: Employs a precise weight-editing technique rather than fine-tuning, which helps avoid degradation of coherence often seen when fine-tuning on top of RLHF'd refusals.
  • Base Model Integrity: Largely preserves the Qwen3-8B's original capabilities, including its 32k context length and dual-mode architecture.

Ideal Use Cases

  • Local Agents: Suitable for applications requiring less restrictive AI behavior.
  • Roleplay: Excels in scenarios where creative freedom and non-refusal are paramount.
  • Research: Valuable for studying refusal mechanisms and model behavior in reasoning-capable LLMs.
  • Resource-Friendly: GGUF quantizations are provided, making it runnable on various consumer GPUs (e.g., RTX 3090/4090 with Q8_0, RTX 4060 with Q4_K_M).