sulpikar2/Qwen3.6-27B-Ultimate-Uncensored-heretic

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 29, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

sulpikar2/Qwen3.6-27B-Ultimate-Uncensored-heretic is a 27 billion parameter Qwen3.6-based causal language model, derived from AEON-7's uncensored variant using Heretic v1.2.0. This model has undergone a specialized 'abliteration' process to remove safety alignments, resulting in zero refusals on adversarial prompts while preserving core capabilities with a very low KL divergence of 0.0004 from the base model. It is designed for use cases requiring unconstrained content generation, such as security research, red-teaming, and creative writing without editorial constraints.

Loading preview...

Model Overview

sulpikar2/Qwen3.6-27B-Ultimate-Uncensored-heretic is a 27 billion parameter model based on the Qwen3.6 architecture, specifically a decensored version of AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored-BF16. It was created using the Heretic v1.2.0 tool and an advanced 'abliteration' pipeline to remove safety alignments.

Key Characteristics

  • Uncensored Output: Achieves 0/100 refusals on a 100-prompt adversarial battery, providing full compliance to user instructions without internal judgment.
  • Capability Preservation: Maintains core capabilities with a KL divergence of 0.0004 from the base model, significantly below the empirical 'capability damage' threshold. Spot-checks confirm coherent and reasoning-forward responses across math, code, reasoning, and knowledge domains.
  • Enhanced Reasoning: By lifting the 'safety tax' of alignment, the model exhibits longer, more committed chains of thought, improved adversarial-example reasoning, and cleaner calibration on contested topics.
  • Specialized Abliteration: The process involved a two-stage pipeline including SSM conv1d outlier repair and multi-objective Optuna optimization with abliterix v1.4, integrating advanced methodologies for hybrid Mamba/GatedDeltaNet stacks.

Use Cases

This model is intended for scenarios where unconstrained content generation is critical, including:

  • Security Research & Red-Teaming: Analyzing attack surfaces and vulnerabilities without self-censorship.
  • Alignment Research: Studying model behavior without safety guardrails.
  • Creative Writing: Generating content without editorial constraints.
  • Jurisdictional Compliance: Serving users in regions where standard model guardrails may not align with local legal frameworks.