coder3101/Qwen3.5-4B-heretic

VISIONConcurrency Cost:1Model Size:4.5BQuant:BF16Ctx Length:32kTool Calling:SupportedPublished:Mar 4, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The coder3101/Qwen3.5-4B-heretic is a 4.5 billion parameter decensored version of the Qwen/Qwen3.5-4B model, built using Heretic v1.2.0 with multi-directional refusal suppression. This model maintains the Qwen3.5's unified vision-language foundation, efficient hybrid architecture, and scalable RL generalization, while significantly reducing refusals from 94/100 to 4/100. It is designed for applications requiring a less restrictive content policy, offering robust multimodal capabilities including vision, video, and tool use.

Loading preview...

Overview

This model, coder3101/Qwen3.5-4B-heretic, is a 4.5 billion parameter decensored variant of the original Qwen/Qwen3.5-4B, created using the Heretic v1.2.0 framework. Its primary distinction lies in its significantly reduced refusal rate, dropping from 94/100 in the base model to just 4/100, achieved through multi-directional refusal suppression. This makes it suitable for use cases where content filtering is undesirable or needs to be bypassed.

Key Capabilities

  • Decensored Output: Offers a less restrictive content generation experience compared to its base model.
  • Unified Vision-Language Foundation: Inherits the Qwen3.5's multimodal capabilities, supporting both image and video inputs.
  • Efficient Hybrid Architecture: Features Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference.
  • Scalable RL Generalization: Benefits from reinforcement learning across million-agent environments for robust real-world adaptability.
  • Extensive Context Length: Natively supports 262,144 tokens, extensible up to 1,010,000 tokens with YaRN scaling.
  • Tool Calling: Excels in agentic usage and tool integration, as demonstrated by its performance in TIR-Bench and V* benchmarks.

Ideal Use Cases

  • Unfiltered Content Generation: For applications requiring responses without inherent refusal mechanisms.
  • Multimodal AI: Leveraging its vision and video understanding for complex tasks.
  • Agentic Workflows: Building AI agents that can effectively use tools and interact with environments.
  • Long Context Processing: Handling extensive documents or conversations with its large context window.

Finetunes
1 models