p-e-w/Qwen3.5-4B-heretic

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:May 5, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

The p-e-w/Qwen3.5-4B-heretic is a 4.5 billion parameter causal language model, based on the Qwen3.5-4B architecture, that has been decensored using the Heretic v1.3.0 tool. This model retains the multimodal capabilities of the original Qwen3.5, including unified vision-language foundation and support for 201 languages, while significantly reducing refusal rates from 93/100 to 40/100. It is designed for applications requiring a powerful, versatile, and less restrictive multimodal AI with a native context length of 262,144 tokens, extensible up to 1,010,000 tokens.

Loading preview...

Model Overview

p-e-w/Qwen3.5-4B-heretic is a 4.5 billion parameter multimodal causal language model derived from the Qwen/Qwen3.5-4B base model, with a key modification: it has been decensored using Heretic v1.3.0. This process significantly alters the model's refusal behavior, reducing it from 93 out of 100 instances in the original model to just 40 out of 100, as measured by KL divergence of 0.0171.

Key Capabilities

  • Decensored Output: Offers a less restrictive response generation compared to its base model, making it suitable for use cases where content filtering is undesirable.
  • Unified Vision-Language Foundation: Inherits Qwen3.5's ability to process and understand both visual and textual inputs, achieving strong performance across reasoning, coding, agents, and visual understanding benchmarks.
  • Efficient Hybrid Architecture: Utilizes Gated Delta Networks and sparse Mixture-of-Experts for high-throughput inference with minimal latency.
  • Scalable RL Generalization: Benefits from reinforcement learning across million-agent environments for robust real-world adaptability.
  • Extensive Multilingual Support: Supports 201 languages and dialects, ensuring broad global applicability.
  • Ultra-Long Context: Natively handles up to 262,144 tokens, extensible to 1,010,000 tokens using YaRN scaling techniques, ideal for complex, long-form tasks.

Good For

  • Applications requiring unfiltered responses: Ideal for research or creative tasks where the base model's content restrictions might be prohibitive.
  • Multimodal AI development: Leveraging its vision-language capabilities for tasks involving image, video, and text understanding.
  • Global deployments: Its extensive language support makes it suitable for international applications.
  • Complex, long-context tasks: Excels in scenarios demanding deep understanding and generation over very long inputs, such as document analysis or extended conversations.