Iwaku-Real/Qwen3-0.6B-Base-heretic-test

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:0.8BQuant:BF16Ctx Length:32kPublished:May 30, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

Iwaku-Real/Qwen3-0.6B-Base-heretic-test is a 0.8 billion parameter causal language model, based on the Qwen3-0.6B-Base architecture developed by Qwen, with a 32,768 token context length. This model has been decensored using the Heretic v1.3.0 tool, specifically modified to reduce refusals compared to its original base model. It is designed for applications requiring a less restrictive language model while maintaining the foundational capabilities of the Qwen3 series.

Loading preview...

Iwaku-Real/Qwen3-0.6B-Base-heretic-test: Decensored Qwen3 Base Model

This model is a modified version of the Qwen/Qwen3-0.6B-Base, a 0.6 billion parameter causal language model from the Qwen3 series, featuring a 32,768 token context length. It has been processed using the Heretic v1.3.0 tool to achieve a decensored output, significantly reducing refusal rates compared to the original model.

Key Differentiators & Capabilities

  • Decensored Output: Modified to produce less restrictive responses, with a reported 2/100 refusal rate compared to 10/100 for the original Qwen3-0.6B-Base.
  • Reproducible Modification: The abliteration process used by Heretic is documented and reproducible, allowing for transparency in its modification.
  • Qwen3 Foundation: Inherits the architectural advancements of the Qwen3 series, including an expanded 36 trillion token pre-training corpus covering 119 languages, refined training techniques, and a three-stage pre-training approach for broad language modeling, reasoning, and long-context comprehension.
  • Optimized for Stability: Incorporates architectural refinements like qk layernorm for improved stability and performance.

When to Consider This Model

  • Applications requiring less restrictive content generation: Ideal for use cases where the default censorship of base models is too limiting.
  • Exploration of decensored LLM behavior: Useful for researchers or developers interested in the effects of abliteration on model responses.
  • Small-scale deployments: Its 0.6 billion parameter size makes it suitable for environments with limited computational resources, while still offering a substantial context window.