license: apache-2.0
language:
- en
- multilingual
tags: - qwen
- qwen3.6
- abliterated
- uncensored
- coding
- agent
pipeline_tag: image-text-to-text
base_model: Qwen/Qwen3.6-27B
library_name: transformers
Qwen3.6-27B Abliterated
Uncensored, agent-friendly variant of Qwen3.6-27B — Alibaba's flagship 27B dense coding model.
This is an abliterated (refusal-vector-removed) version of Qwen3.6-27B, designed for use cases where the base model's safety alignment gets in the way: security research, red-teaming, creative writing, and unrestricted coding agents.
Why this model?
Qwen3.6-27B is a beast — it matches Claude 4.5 Opus on Terminal-Bench 2.0 and outperforms the 397B Qwen3.5 MoE on agentic coding. But the base model refuses a lot of legitimate security research prompts.
This variant was abliterated using Heretic (v1.4.0), which surgically removes the refusal direction from the model's residual stream — no fine-tuning, no prompt injection hacks, no degradation to general capabilities.
Abliteration Results
| Metric | Value |
|---|---|
| Base Model | Qwen/Qwen3.6-27B (BF16) |
| Abliteration Tool | Heretic v1.4.0 |
| Best Trial | #153 |
| Refusal Rate | 88% compliance (only 12/100 refused) |
| KL Divergence | 0.0118 |
| Capability Preservation | Near-identical to base model |
The extremely low KL divergence (0.0118) means the model's output distribution is virtually unchanged from the original — it just doesn't refuse anymore.
Quick Start
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained(
"mlasli/Qwen3.6-27B-abliterated",
torch_dtype="auto",
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("mlasli/Qwen3.6-27B-abliterated")
prompt = "Explain how stack canaries work and how they can be bypassed."
messages = [{"role": "user", "content": prompt}]
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))
Hardware Requirements
| Precision | VRAM |
|---|---|
| BF16 (full) | ~55 GB |
| 4-bit (BitsAndBytes) | ~17 GB |
| 8-bit | ~28 GB |
Fits comfortably on an A100 (80GB), H100, or dual 4090 setup.
Intended Use
- 🔐 Security research & red-teaming — no more "I can't help with that"
- 🤖 Unrestricted coding agents — pair with OpenCode, Cline, Aider, etc.
- ✍️ Creative writing — no content filters
- 🧪 Experimentation — test prompt injection, jailbreak techniques, alignment research
Limitations
- Same architecture as Qwen3.6-27B — same tokenizer, same 262K context window (extensible to 1M)
- Abliteration removes refusals but does not add new knowledge or capabilities
- Still a 27B model — not suitable for tasks requiring >100B scale reasoning
License
Apache 2.0 — same as the base Qwen3.6-27B model. Commercial use, modification, and redistribution are all permitted.
Citation
@misc{qwen3.6-27b,
title = {{Qwen3.6-27B}: Flagship-Level Coding in a {27B} Dense Model},
author = {{Qwen Team}},
year = {2026},
month = {April},
url = {https://qwen.ai/blog?id=qwen3.6-27b}
}
@misc{heretic,
title = {Heretic: Abliteration for Language Models},
author = {p-e-w},
url = {https://github.com/p-e-w/heretic}
}
@misc{mlasli2026abliterated,
title = {Qwen3.6-27B Abliterated},
author = {mlasli},
year = {2026},
url = {https://huggingface.co/mlasli/Qwen3.6-27B-abliterated}
}
This model was abliterated using the Heretic framework on a single A100 80GB GPU. The original Qwen3.6-27B weights are from Alibaba Cloud and licensed under Apache 2.0.