joyfox/Qwen3.8-27B-Uncensored-JoyFox-Aggressive

Hugging Face
VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

The joyfox/Qwen3.8-27B-Uncensored-JoyFox-Aggressive model is a 27 billion parameter variant of the Qwen3.8 architecture, developed by JoyFox. This model is specifically fine-tuned to reduce over-refusal and provide direct, substantive responses to legitimate requests, while preserving its native multimodal capabilities including vision and Multi-Token Prediction (MTP). It is designed for applications requiring a less restrictive and more direct conversational style, maintaining the original Qwen3.8 architecture, tokenizer, and vision stack.

Loading preview...

Overview

joyfox/Qwen3.8-27B-Uncensored-JoyFox-Aggressive is a 27 billion parameter model based on the Qwen3.8 architecture, developed by JoyFox. Its primary distinction is a significant reduction in over-refusal and policy-style preambles, aiming for direct and substantive responses to a wide range of legitimate requests. This "Aggressive" behavior means it is less likely to refuse sensitive or ambiguous prompts, making it suitable for use cases where a more unconstrained response style is desired.

Key Capabilities

  • Reduced Over-Refusal: Evaluated with a 0.0% benign refusal rate compared to 59.2% for the official Qwen3.8-27B, ensuring more direct answers.
  • Multimodal Support: Retains the full vision stack and multimodal projector from the base Qwen3.8 model, allowing for image input and understanding.
  • Native MTP Preservation: Includes native Multi-Token Prediction (MTP) weights, supporting speculative decoding for potentially faster inference.
  • Architectural Integrity: Preserves the official tokenizer, chat template, and configuration, ensuring compatibility with the base model's ecosystem.

Good For

  • Applications requiring a direct and less restrictive AI assistant.
  • Use cases where the model needs to engage with sensitive or ambiguous topics without excessive refusal.
  • Multimodal tasks that benefit from both text and image understanding.
  • Developers seeking a Qwen3.8-based model with enhanced conversational directness for benign, fictional, analytical, or educational purposes.