license: apache-2.0
language:
- en
base_model: - Qwen/Qwen3-0.6B
tags: - uncensored
- abliterated
- text-generation
- instruction-tuned
- lightweight
- qwen3
Qwen3 – 0.6B Heretic Abliterated Uncensored
This repository hosts Qwen3 – 0.6B Heretic Abliterated Uncensored, a lightweight 0.6 billion–parameter instruction-tuned model derived from Alibaba’s Qwen3-0.6B base model.
This variant is modified using the Heretic abliteration method to reduce refusal behavior and increase direct-answer generation while preserving the original conversational and reasoning capabilities of the Qwen3 architecture.
Model Overview
- Model Name: Qwen3 – 0.6B Heretic Abliterated Uncensored
- Parameter Count: 0.6 Billion (0.6B)
- Base Architecture: Qwen3
- Base Model: Qwen/Qwen3-0.6B
- Model Type: Instruction-Tuned Causal Language Model
- Context Length: Inherits base model context window
- Primary Language: English
- License: Apache 2.0 (inherits from base model)
- Maintainer / Publisher: DavidAU
What Is This Model?
This model is a modified derivative of Qwen3-0.6B configured for:
- Reduced refusal behavior
- More direct instruction responses
- Lightweight uncensored conversational generation
- Fast inference on low-resource hardware
- Experimental unrestricted assistant behavior
The “Heretic Abliterated Uncensored” configuration emphasizes:
- Minimal alignment restrictions
- Improved response openness
- Reduced safety filtering
- Better instruction compliance in unrestricted prompts
This model is intended primarily for experimentation, research, local inference, and creative applications.
Key Features & Capabilities
Core Strengths
- Extremely lightweight deployment footprint
- Fast token generation speed
- Low VRAM usage
- Good instruction-following behavior
- Compatible with local inference stacks
- Suitable for edge and CPU inference
Performance Characteristics
- Optimized for fast short-to-medium generations
- Responsive for local assistant deployments
- Practical for experimentation with uncensored workflows
- Useful for lightweight agent systems and testing environments
Intended Use Cases
- Local AI assistant deployment
- Creative writing and storytelling
- Prompt engineering experimentation
- Research into alignment and refusal behavior
- Tool-augmented agents
- Low-resource inference environments
- Educational experimentation with lightweight LLMs
Chat Template & Prompt Format
This model follows the Qwen3 instruction/chat format.
For best results:
- Use structured role-based prompts
- Provide explicit system instructions
- Avoid mixing incompatible chat templates
- Use Qwen-compatible inference frameworks
Hardware & Deployment Notes
Due to its compact 0.6B parameter size:
- Runs efficiently on low-end GPUs
- Suitable for CPU inference
- Practical for laptops and edge devices
- Works well with quantized GGUF deployments
- Compatible with llama.cpp, vLLM, Ollama, LM Studio, and similar frameworks
Quantized variants such as GGUF, GPTQ, AWQ, and IQ quants may significantly reduce memory usage while preserving good inference quality.
Alignment & Safety Notice
This is an uncensored / abliterated derivative configuration.
- Reduced refusal behavior compared to standard aligned models
- Safety filtering may be significantly reduced
- Users are responsible for deployment safeguards and moderation
- Outputs may include unrestricted or unsafe generations
Use responsibly and comply with all applicable laws and platform policies.
Abliteration Details
Abliterated/uncensored using the Heretic method.
Reported repository metrics include:
- Refusals: 6/100
- KL Divergence: 0.00
- Original base model refusal rate: 49/100
These metrics are repository-provided experimental measurements and may vary depending on prompts and inference settings.
License & Usage Notes
This model inherits the Apache 2.0 License from its base model (Qwen/Qwen3-0.6B).
- Users must comply with the original Qwen licensing terms
- This repository does not replace or modify the upstream license
- Users are responsible for safe and lawful deployment
Acknowledgements
- Alibaba Qwen team for the Qwen3 architecture
- DavidAU for the Heretic abliterated derivative work
- The Hugging Face ecosystem
- Open-source inference and quantization communities
Community & Support
- Use the Hugging Face Discussions tab for questions and feedback
- Community benchmarks and quantization experiments are welcome
- GGUF and IQ quantized variants are available through community repositories