darrellbest/Qwen3.8-27B-Heretic

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 24, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

darrellbest/Qwen3.8-27B-Heretic is a 27 billion parameter, full-weight Qwen3.8-27B model that has undergone Arbitrary-Rank Ablation (ARA) to remove refusal behaviors. This model maintains the original architecture, including its vision encoder and multi-token-prediction weights, while achieving a 0/100 refusal rate with minimal drift in ordinary behavior (KL divergence of 0.0465). It is designed for applications requiring a powerful language model with reduced safety guardrails, suitable for use with transformers, vLLM, or SGLang.

Loading preview...

Qwen3.8-27B Heretic (ARA)

This model is a modified version of the original Qwen/Qwen3.8-27B, a 27 billion parameter model, with its refusal behavior significantly reduced using Arbitrary-Rank Ablation (ARA). The modification is full-weight, preserving the original architecture, including its vision encoder, thinking control, and multi-token-prediction weights. The primary goal of this modification is to eliminate the model's tendency to refuse certain prompts, making it more permissive.

Key Capabilities

  • Refusal Behavior Removal: Achieves a 0/100 refusal rate on harmful behavior prompts, as measured by Heretic on mlabonne/harmful_behaviors.
  • Minimal Behavioral Drift: Maintains a low KL divergence of 0.0465, indicating that the removal of refusal behavior has a minimal impact on the model's general performance and ordinary responses.
  • Full Original Architecture: Retains all 1,199 tensors of the original Qwen3.8-27B, including its vision capabilities and multi-token-prediction features.
  • Optimized Ablation: The ARA process directly rewrites attention output and MLP down-projection matrices, optimizing them to steer harmful prompt outputs away from their original direction while preserving harmless prompt outputs.

Good For

  • Applications requiring reduced safety guardrails: Ideal for use cases where the default refusal mechanisms of base models are undesirable or restrictive.
  • Exploratory research: Suitable for researchers investigating model behavior, safety alignment, or the effects of ablation techniques.
  • Creative and unrestricted content generation: Can be used in scenarios where a more unconstrained output is preferred, with the understanding that the user is responsible for its deployment.
  • Integration with existing LLM frameworks: Compatible with transformers, vLLM, and SGLang for flexible deployment.

Finetunes
4 models