martin-t-studio/Qwen3-4B-Instruct-2507-heretic
TEXT GENERATIONConcurrency Cost:1Model Size:4BQuant:BF16Ctx Length:32kPublished:Apr 7, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

The martin-t-studio/Qwen3-4B-Instruct-2507-heretic is a 4 billion parameter instruction-tuned causal language model, derived from the Qwen3-4B-Instruct-2507 base model. Developed by Martin T. Studio using the Heretic v1.2.0 tool, this version is specifically decensored, featuring significantly reduced safety filtering compared to its original counterpart. It is optimized for research, creative writing, and experimental purposes where the generation of controversial or sensitive content is intentionally permitted.

Loading preview...

Overview

This model, martin-t-studio/Qwen3-4B-Instruct-2507-heretic, is a decensored variant of the Qwen/Qwen3-4B-Instruct-2507 base model. Created by Martin T. Studio using the Heretic v1.2.0 tool, its primary distinction is the intentional removal of most alignment-based refusal mechanisms.

Key Characteristics & Performance

  • Decensored Nature: Unlike the original model which had 100/100 refusals, this "heretic" version exhibits only 16/100 refusals, indicating a significantly reduced safety filter.
  • Abliteration Parameters: Specific parameters like direction_index, attn.o_proj.max_weight, and mlp.down_proj.max_weight were adjusted during the decensoring process.
  • KL Divergence: The model shows a KL divergence of 0.0929 compared to the original, reflecting the changes introduced.

Intended Use Cases

This model is designed for specific applications where unfiltered content generation is desired:

  • Research: Exploring the capabilities and behaviors of large language models without typical safety constraints.
  • Creative Writing: Generating diverse and unrestricted content for fictional narratives, roleplay, or artistic expression.
  • Experimental Purposes: Testing boundaries and understanding the implications of decensored AI outputs.

Important Considerations

Users must be aware of the reduced safety filtering. The model may produce controversial, sensitive, or NSFW content. Users are solely responsible for their prompts and the resulting outputs, and usage should comply with local laws and ethical standards. It is not recommended for public-facing commercial applications or use by minors.

Finetunes
2 models