richardyoung/ deepseek-r1-0528-qwen3-8b-heretic:Q5_K_M

191 2 days ago

Uncensored DeepSeek-R1-0528-Qwen3-8B via Heretic: 4/100 refusals (from 89), KL 0.041. Reasoning model, MIT. 

ollama run richardyoung/deepseek-r1-0528-qwen3-8b-heretic:Q5_K_M

Details

2 days ago

1c1d59bff4ae · 5.9GB

model
qwen3
·
8.19B
·
Q5_K_M

Readme

DeepSeek-R1-0528-Qwen3-8B-Heretic

Uncensored DeepSeek-R1-0528-Qwen3-8B via Heretic: 4⁄100 refusals (from 89), KL 0.041. Reasoning model, MIT.

🚀 Overview

This is an abliterated build of deepseek-ai/DeepSeek-R1-0528-Qwen3-8B, DeepSeek’s R1-0528 reasoning distilled into Qwen3-8B: a reasoning model that emits chains before answering. Refusal behavior was reduced using the Heretic library with KL-targeted parameters that preserve the model’s coherence.

📊 Abliteration Results

Metric Before After
Refusals 89⁄100 4⁄100
Reduction – 96%
KL Divergence – 0.041

The very low KL divergence (0.041, far below the 0.5 “damage” threshold) means the model retains essentially all of its original capabilities and coherence.

🎯 Key Features

  • Reduced censorship: 96% fewer refusals on typical “unsafe” prompts
  • Near-zero quality loss: KL 0.041
  • Reasoning model ( chains)
  • R1-0528 reasoning distilled into an 8B model
  • 128K-token context
  • MIT license
  • Reproducible: full reproduction data on Hugging Face

🏷️ Available Versions

Tag Size BPW Notes
latest / Q4_K_M 5.0 GB 4.85 Recommended
Q5_K_M 5.9 GB 5.68 Higher quality
Q6_K 6.7 GB 6.56 Very high quality
Q8_0 8.7 GB 8.5 Near-lossless

💻 Quick Start

ollama run richardyoung/deepseek-r1-0528-qwen3-8b-heretic           # recommended (Q4_K_M)
ollama run richardyoung/deepseek-r1-0528-qwen3-8b-heretic:Q8_0      # near-lossless

🛠️ Use Cases

  • Math, coding and logic with visible reasoning
  • Research and red-teaming on reasoning models
  • Education

📋 System Requirements

VRAM Recommended tier
6 GB+ Q4_K_M (5.0 GB file)
10 GB+ Q8_0 (8.7 GB file)

🔧 Technical Details

  • Base Model: deepseek-ai/DeepSeek-R1-0528-Qwen3-8B
  • Parameters: 8.2B (qwen3 architecture)
  • Context Length: 128K tokens
  • Quantization: GGUF via llama.cpp (text generation)
  • Abliteration: Heretic v2.0.0.dev0 by p-e-w (Trial 173: 89→4 refusals @ KL 0.041)
  • License: MIT (inherited from the base model)

⚠️ Disclaimer

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

🙏 Acknowledgments

  • Base Model: DeepSeek
  • Abliteration: Heretic by p-e-w
  • Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro