Dotalix/Qwen2.5-3B-hereticc

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:3.1BQuant:BF16Ctx Length:32kPublished:Mar 5, 2026License:qwen-researchArchitecture:Transformer Warm

Dotalix/Qwen2.5-3B-hereticc is a 3.1 billion parameter causal language model, based on the Qwen2.5 architecture, specifically modified using Heretic v1.2.0 to be a decensored version of the original Qwen/Qwen2.5-3B. This model retains the Qwen2.5 improvements in coding, mathematics, instruction following, and long text generation, while being optimized for uncensored and decensored text generation. It supports a 32,768 token context length and over 29 languages, making it suitable for applications requiring less restrictive content filtering.

Loading preview...

Dotalix/Qwen2.5-3B-hereticc: Decensored Qwen2.5-3B

This model is a decensored version of the Qwen/Qwen2.5-3B base model, created using the Heretic v1.2.0 tool. It aims to provide a less restrictive language generation experience compared to its original counterpart, as indicated by its lower refusal rate (2/100 compared to 3/100 for the original).

Key Capabilities Inherited from Qwen2.5-3B

  • Enhanced Knowledge & Reasoning: Significant improvements in coding and mathematics, leveraging specialized expert models.
  • Instruction Following: Better adherence to instructions and more resilient to diverse system prompts, aiding in role-play and chatbot implementations.
  • Long Text Generation: Improved capability for generating texts over 8K tokens.
  • Structured Data Handling: Better understanding of structured data like tables and improved generation of structured outputs, especially JSON.
  • Multilingual Support: Supports over 29 languages, including English, Chinese, French, Spanish, German, and more.
  • Long Context: Features a full 32,768 token context length.

Unique Characteristics

  • Decensored Output: Modified to reduce content refusals, making it suitable for use cases where less content filtering is desired.
  • Abliteration Parameters: Specific parameters were adjusted during the Heretic modification process to achieve its decensored nature, affecting attention and MLP projections.

Good for

  • Applications requiring a 3.1 billion parameter model with reduced content restrictions.
  • Tasks involving coding, mathematics, and complex instruction following where the original Qwen2.5-3B's capabilities are beneficial, but with a preference for uncensored outputs.
  • Generating long, structured, or multilingual texts without the typical refusal mechanisms of standard models.

Finetunes
3 models