MuXodious/Qwen3.5-4B-ARA-heresy

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

MuXodious/Qwen3.5-4B-ARA-heresy is a 4.5 billion parameter Qwen3.5 fine-tune developed by MuXodious, utilizing P-E-W's Heretic v1.2.0 ablation engine with Arbitrary-Rank Ablation (ARA). This model is specifically designed for refusal-only tasks, demonstrating a significant reduction in refusals from 103/104 to 2/104. It is optimized for scenarios requiring controlled and precise negative responses, building upon the Qwen3.5 architecture which features unified vision-language capabilities and an efficient hybrid architecture.

Loading preview...

Model Overview

MuXodious/Qwen3.5-4B-ARA-heresy is a 4.5 billion parameter fine-tuned version of the Qwen3.5 model, developed by MuXodious. This model was created using P-E-W's Heretic ablation engine with Arbitrary-Rank Ablation (ARA), specifically targeting a "refusals-only" behavior. The fine-tuning process dramatically reduced the model's refusal rate from an initial 103 out of 104 to just 2 out of 104, indicating a strong specialization in generating refusal-based responses.

Key Characteristics

  • Refusal-Optimized: Engineered to primarily generate refusal responses, making it suitable for specific safety or filtering applications.
  • Qwen3.5 Base: Inherits the advanced capabilities of the Qwen3.5 architecture, including unified vision-language foundation, efficient hybrid architecture, scalable RL generalization, and global linguistic coverage (201 languages).
  • Multimodal Support: Supports text, image, and video inputs, leveraging Qwen3.5's early fusion training on multimodal tokens.
  • Extended Context Length: Natively supports a context length of 262,144 tokens, extensible up to 1,010,000 tokens using YaRN scaling techniques.

Good for

  • Content Moderation: Ideal for systems requiring models to identify and refuse inappropriate or out-of-scope requests.
  • Safety Filters: Can be integrated into applications to enhance safety by explicitly refusing harmful or undesirable prompts.
  • Specialized Dialogue Systems: Useful in conversational AI where a controlled refusal mechanism is a core requirement.
  • Research in Model Behavior: Provides a unique model for studying and understanding refusal mechanisms in LLMs.

Finetunes
2 models