Youssofal/Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Youssofal/Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16 is a 27.8 billion parameter vision-language model based on Qwen's Qwen3.6-27B, with a 32768 token context length. This model has undergone a Heretic-style ablation process to significantly reduce refusal behaviors while preserving its full multimodal capabilities for image and video input. It is designed for applications requiring a powerful, uncensored multimodal LLM that can handle a broader range of prompts than its base model.

Loading preview...

Model Overview

Youssofal/Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16 is a 27.8 billion parameter, BF16 vision-language model derived from Qwen's Qwen3.6-27B. Its core differentiator is the application of a Heretic-style two-stage MPOA pipeline (Magnitude Preservation Orthogonalization Ablation) to attenuate refusal behaviors at the weight level. This process significantly reduces the model's tendency to refuse prompts, as evidenced by hand-read refusal checks showing a reduction from 25/25 refusals to 0/25 direct passes (with some deflections) under a jailbreak system prompt, and 16/25 deflections without a system prompt.

Key Capabilities

  • Reduced Refusal Behavior: Engineered to answer prompts that the original Qwen3.6-27B would typically refuse, offering greater flexibility.
  • Full Multimodal Support: Retains the complete vision and video multimodal architecture of the base Qwen3.6-27B, allowing for image and video input processing.
  • High Fidelity to Base Model: Achieves low distributional divergence (KL 0.0282) compared to the base model, ensuring general capabilities remain largely intact.
  • Robust Architecture: Features 64 text layers and a hybrid attention mechanism, supporting a 32768 token context length.

Use Cases

This model is suitable for applications where an uncensored, multimodal large language model is required, particularly for tasks involving image and video understanding, and scenarios where the base model's refusal patterns would be restrictive. Users should be aware of the attenuated refusal behavior and are responsible for its deployment.