Qwen3-8B-Abliterated

Abliterated (uncensored) version of Qwen3-8B for unrestricted conversations, reasoning, and creative writing.

🚀 Overview

This model is an abliterated version of Qwen/Qwen3-8B, with refusal behavior reduced through targeted weight modification. The abliteration process uses the Heretic library with conservative parameters to preserve model coherence while removing restrictions. It retains the full Qwen3 feature set, including seamless switching between thinking mode (for complex reasoning, math, and code) and non-thinking mode (for efficient, general-purpose dialogue).

📊 Abliteration Results

Refusal metrics pending re-measurement.

A low KL divergence (< 1.0) indicates the model maintains its original capabilities and coherence.

🎯 Key Features

• Reduced Censorship: Fewer refusals on typical “unsafe” prompts via conservative abliteration
• Preserved Quality: Conservative weight modification maintains model coherence and reasoning ability
• Thinking Mode: Retains Qwen3 dual-mode reasoning, switch with /think and /no_think
• Multilingual & Coding: Full Qwen3 capabilities including 100+ languages, coding, and math
• GGUF Quants: Multiple quantization levels for flexible deployment via Ollama and llama.cpp

🏷️ Available Versions

Only the Q4_K_M build is currently published. The BPW guide below shows where additional quants sit if released:

Quant	BPW	Profile
IQ3_M	3.66	Smallest, for low VRAM
IQ4_XS	4.25	Great quality/size balance
Q4_K_M	4.85	Recommended
Q5_K_M	5.68	Higher quality
Q6_K	6.56	Very high quality
Q8_0	8.5	Near-lossless

💻 Quick Start

ollama run richardyoung/qwen3-8b-abliterated
ollama run richardyoung/qwen3-8b-abliterated:latest

🛠️ Use Cases

• Creative Writing: Generate unrestricted fiction, scenarios, and roleplay
• Reasoning & Research: Explore topics and work through problems without artificial limitations
• Red Teaming: Test AI safety without built-in guardrails
• Coding & Math: Use thinking mode for complex code generation and step-by-step math
• Education: Learn about topics that are typically refused

📋 System Requirements

🔧 Technical Details

• Base Model: Qwen/Qwen3-8B
• Parameters: 8.2B (6.95B non-embedding), 36 layers
• Context Length: 32,768 tokens native; up to 131,072 (128K) with YaRN
• Quantization: Q4_K_M (4.85 bits per weight)
• Abliteration: Heretic by p-e-w, with conservative weight modification

⚠️ Disclaimer

This model has reduced safety guardrails. The reduction of refusal behavior means the model will engage with a wider range of prompts. Use responsibly and in accordance with applicable laws and regulations.

🙏 Acknowledgments

• Base Model: Alibaba / Qwen team for Qwen3-8B
• Abliteration: Heretic by p-e-w
• Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro