MuXodious/Qwen3.5-27B-tainted-heresy

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Feb 28, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

MuXodious/Qwen3.5-27B-tainted-heresy is a 27 billion parameter Qwen3.5 fine-tune, developed by MuXodious using P-E-W's Heretic ablation engine. This model is specifically engineered to reduce model-unique refusals and overt non-compliance, achieving a refusal count of 13/104 with a KL Divergence of 0.1109. It maintains the Qwen3.5 base's unified vision-language foundation, efficient hybrid architecture, and scalable RL generalization, making it suitable for applications requiring reduced model guardrails while retaining strong multimodal and reasoning capabilities.

Loading preview...

Overview

MuXodious/Qwen3.5-27B-tainted-heresy is a 27 billion parameter model derived from the Qwen3.5 base, fine-tuned using P-E-W's Heretic ablation engine. The primary objective of this fine-tuning was to significantly reduce model-unique refusals and overt non-compliance, achieving a notable reduction from 104/104 initial refusals to 13/104, with a KL Divergence of 0.1109. This process, termed "heretication," aims to modify the model's behavior regarding disclaimers and non-compliance while preserving its core capabilities.

Key Capabilities

  • Reduced Refusals: Specifically engineered to minimize model-unique refusals and overt non-compliance, making it more permissive in its responses.
  • Unified Vision-Language Foundation: Inherits Qwen3.5's ability for early fusion training on multimodal tokens, excelling in reasoning, coding, agents, and visual understanding.
  • Efficient Hybrid Architecture: Utilizes Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference with optimized latency and cost.
  • Scalable RL Generalization: Benefits from reinforcement learning scaled across million-agent environments for robust real-world adaptability.
  • Multilingual Support: Expanded support for 201 languages and dialects, ensuring broad global applicability.
  • Long Context Window: Features a native context length of 262,144 tokens, extensible up to 1,010,000 tokens, crucial for complex tasks.

Good For

  • Applications requiring less restrictive model behavior: Ideal for use cases where the default Qwen3.5 model's refusal mechanisms are too stringent.
  • Multimodal tasks: Strong performance in vision-language understanding, including STEM, puzzle-solving, and general VQA.
  • Agentic workflows: Excels in general agent, search agent, and visual agent benchmarks, making it suitable for tool-calling and automated task execution.
  • Long-context processing: Capable of handling extensive inputs for tasks like document understanding and complex reasoning.

Finetunes
2 models