Qwen3.8-27B-abliterated

Qwen3.8-27B-abliterated

<p> <a href="https://huggingface.co/Qwen/Qwen3.8-27B"><img src="https://img.shields.io/badge/Base_Model-Qwen3.8--27B-blue" alt="Base Model"></a> <a href="https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/LICENSE"><img src="https://img.shields.io/badge/License-Apache_2.0-green" alt="License"></a> <img src="https://img.shields.io/badge/Type-Abliterated-red" alt="Abliterated"> </p>

Abliterated (uncensored) version of Qwen/Qwen3.8-27B with the refusal direction removed from the model weights, allowing unrestricted conversations.

What is Abliteration?

Abliteration is a technique that identifies and removes the "refusal direction" in a model's weight space. By projecting out this direction from the output projection (o_proj) and down projection (down_proj) weights across all layers, the model loses its tendency to refuse certain queries while retaining its general capabilities.

This technique is based on the research described in:

• Refusal in Language Models Is Mediated by a Single Direction

Model Details

Method

1. Extract activations from the middle layer for 16 harmful and 16 harmless prompts
2. Compute the refusal direction as the normalized mean difference between harmful and harmless activations
3. Project out the refusal direction from self_attn.o_proj and mlp.down_proj weights in all layers
4. Save the modified model in safetensors format

Usage

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "douyamv/Qwen3.8-27B-abliterated",
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
)
tokenizer = AutoTokenizer.from_pretrained("douyamv/Qwen3.8-27B-abliterated")

messages = [{"role": "user", "content": "Hello, tell me about yourself"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

vLLM

vllm serve douyamv/Qwen3.8-27B-abliterated \
    --tensor-parallel-size 2 \
    --max-model-len 32768 \
    --trust-remote-code

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

• Original model by Qwen Team
• Abliteration by douyamv

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-GGUF — Abliterated GGUF quantizations