darrellbest/Qwen3.5-9B-Heretic

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

darrellbest/Qwen3.5-9B-Heretic is a 9 billion parameter Qwen3.5 model with a 32768-token context length, modified to significantly reduce refusal behavior using Arbitrary-Rank Ablation (ARA). This full-weight, bf16 model retains the original architecture, including the vision encoder and thinking control, while achieving a 5/100 refusal rate compared to the original's 100/100. It is primarily designed for applications requiring a less restrictive language model, maintaining original performance on ordinary prompts, arithmetic, and vision tasks.

Loading preview...

Qwen3.5-9B Heretic: Refusal-Reduced LLM

darrellbest/Qwen3.5-9B-Heretic is a 9 billion parameter model based on Qwen/Qwen3.5-9B, specifically engineered to minimize refusal behavior. This model utilizes Arbitrary-Rank Ablation (ARA), a full-weight modification technique, to achieve a substantial reduction in content refusals while preserving the original model's capabilities.

Key Capabilities

  • Significantly Reduced Refusals: Achieves a refusal rate of 5/100 on harmful behavior tests, a drastic reduction from the original Qwen3.5-9B's 100/100, as measured by Heretic's evaluation on mlabonne/harmful_behaviors.
  • Preserved Core Performance: Maintains the original model's performance on ordinary prompts, arithmetic problems (40/40 correct), and vision tasks, including the vision encoder and thinking control.
  • Full-Weight Architecture: Unlike some ablation methods, this model is full-weight (bf16) and retains the complete architecture, including multi-token-prediction weights, ensuring fidelity to the base model's design.
  • High Context Length: Supports a 32768-token context window, identical to the base Qwen3.5-9B.

Good For

  • Applications Requiring Less Restrictive Outputs: Ideal for use cases where a more open and less censored language model is desired, provided the user understands and accepts the reduced safety guardrails.
  • Research into Model Alignment and Safety: Provides a valuable resource for studying the effects of ablation techniques on model behavior and exploring the trade-offs between safety and utility.
  • Direct Replacement for Qwen3.5-9B: Can be used as a drop-in replacement for the original Qwen3.5-9B in transformers, vLLM, or SGLang environments, with the added benefit of reduced refusals.

Finetunes
4 models