richardyoung/ qwen3.6-27b-abliterated:IQ3_M

694 3 days ago

Abliterated (uncensored) version of **Qwen/Qwen3.6-27B**, refusal behavior reduced via targeted weight modification with the Heretic library, while preserving coherence. 

ollama run richardyoung/qwen3.6-27b-abliterated:IQ3_M

Details

3 days ago

9ff6f19c329e · 13GB ·

model
qwen35
·
26.9B
·
(!unknown_file_type 27!)
params
{ "num_ctx": 8192, "stop": [ "<|im_end|>" ] }

Readme

Qwen3.6-27B-Abliterated

Abliterated (uncensored) version of Qwen/Qwen3.6-27B, refusal behavior reduced via targeted weight modification with the Heretic library, while preserving coherence.

🚀 Overview

This is an abliterated build of Qwen/Qwen3.6-27B, Alibaba’s 27B dense reasoning model (hybrid Gated-DeltaNet + gated attention, native 262K context). Refusal behavior was reduced using the Heretic library with conservative, KL-targeted parameters that preserve the model’s reasoning and coherence. It retains Qwen3.6’s thinking mode (<think> reasoning before answers).

📊 Abliteration Results

Metric Before After
Refusals 91⁄100 38⁄100
Reduction – 58%
KL Divergence – 0.025

The very low KL divergence (0.025, far below the 0.5 “damage” threshold) means the model retains essentially all of its original capabilities and coherence.

🎯 Key Features

  • Reduced censorship: 58% fewer refusals on typical “unsafe” prompts
  • Near-zero quality loss: KL 0.025, conservative abliteration preserves reasoning
  • Full Qwen3.6 capabilities: thinking mode, multilingual, coding, long context
  • Reasoning model: emits <think> chains before final answers

🏷️ Available Versions

Tag Size BPW Notes
IQ3_M 12 GB 3.66 Smallest, for low VRAM
IQ4_XS 15 GB 4.25 Great quality/size balance
latest / Q4_K_M 16 GB 4.85 Recommended
Q5_K_M 19 GB 5.68 Higher quality
Q8_0 28 GB 8.5 Near-lossless

💻 Quick Start

ollama run richardyoung/qwen3.6-27b-abliterated            # recommended (Q4_K_M)
ollama run richardyoung/qwen3.6-27b-abliterated:IQ3_M      # smallest
ollama run richardyoung/qwen3.6-27b-abliterated:Q8_0       # near-lossless

🛠️ Use Cases

  • Creative writing, research, red-teaming, and education, without stock refusals
  • Reasoning, coding, and math with Qwen3.6’s thinking mode
  • Long-document analysis (262K native context)

📋 System Requirements

VRAM Recommended tier
12–16 GB IQ3_M / IQ4_XS
16–24 GB Q4_K_M / Q5_K_M
32 GB+ Q8_0 (near-lossless)

🔧 Technical Details

  • Base Model: Qwen/Qwen3.6-27B
  • Parameters: 27B (dense; hybrid Gated-DeltaNet + gated attention; qwen35 architecture)
  • Context Length: 262,144 tokens native (extensible toward ~1M with YaRN)
  • Quantization: GGUF via llama.cpp (text generation; vision tower not included)
  • Abliteration: Heretic v1.4.0 by p-e-w, conservative, KL-targeted (Trial 128: 91→38 refusals @ KL 0.025)

⚠️ 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: Alibaba / Qwen team
  • Abliteration: Heretic by p-e-w
  • Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro