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

Andycurrent/Qwen3-0.6B-heretic-abliterated-uncensored-GGUF

Qwen3 – 0.6B Heretic Abliterated Uncensored