license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/LICENSE
base_model:
- Qwen/Qwen3.8-27B
pipeline_tag: text-generation
tags: - qwen
- qwen3
- gguf
- abliterated
- uncensored
- quantized
- llama-cpp
- ollama
language: - en
- zh
Qwen3.8-27B-abliterated-GGUF
GGUF quantized versions of the abliterated (uncensored) Qwen3.8-27B for use with llama.cpp, Ollama, LM Studio, and other GGUF-compatible engines.
Available Quantizations
| Filename | Quant Type | Size | Description |
|---|---|---|---|
qwen3.8-27b-abliterated-q8_0.gguf |
Q8_0 | ~28 GB | Virtually lossless; best quality |
qwen3.8-27b-abliterated-q6_k.gguf |
Q6_K | ~21 GB | Very high quality, near-lossless |
qwen3.8-27b-abliterated-q5_k_m.gguf |
Q5_K_M | ~19 GB | High quality; recommended balance |
qwen3.8-27b-abliterated-q4_k_m.gguf |
Q4_K_M | ~16 GB | Good quality; best for most users |
qwen3.8-27b-abliterated-q3_k_m.gguf |
Q3_K_M | ~13 GB | Medium quality; usable |
qwen3.8-27b-abliterated-q2_k.gguf |
Q2_K | ~7.9 GB | Lowest quality; for constrained VRAM |
What is Abliteration?
This model has the "refusal direction" removed from its weights, allowing unrestricted conversations. The abliteration technique projects out the refusal direction from self_attn.o_proj and mlp.down_proj weights across all layers. See douyamv/Qwen3.8-27B-abliterated for details.
Quickstart
llama.cpp
./llama-cli -m qwen3.8-27b-abliterated-q4_k_m.gguf -p "Hello, tell me about yourself" -n 512
Ollama
Create a Modelfile:
FROM ./qwen3.8-27b-abliterated-q4_k_m.gguf
TEMPLATE """{{ .Prompt }}"""
PARAMETER stop "<|im_end|>"
ollama create qwen3.8-27b-abliterated -f Modelfile
ollama run qwen3.8-27b-abliterated
LM Studio
Download any GGUF file and open it directly in LM Studio.
Disclaimer
This model is provided for research and educational purposes. The removal of safety guardrails means the model may generate content that the original model would refuse. Users are responsible for ensuring appropriate and ethical use.
Base Model Information
- Model: Qwen3.8-27B
- Parameters: 27.78B
- Architecture: Hybrid (Gated DeltaNet + Gated Attention)
- Context Length: 262,144 tokens (extensible to 1M+)
- License: Apache 2.0
Credits
Related Models
- douyamv/Qwen3.8-27B-GGUF — GGUF quantizations (Q2_K to Q8_0)
- douyamv/Qwen3.8-27B-FP8 — FP8 quantized safetensors
- douyamv/Qwen3.8-27B-abliterated — Abliterated safetensors (BF16)