Qwen3.8-27B-abliterated
27BQwen
from Qwen/Qwen3.8-27B
↓ 3.9M♥ 1.7K166 repos
Qwen3.8-27B-Coder390-EfficientThink-abliterated
Qwen3.8-27B-Coder390-EfficientThink-abliterated is a refusal-ablated variant of nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2, published on Hugging Face by huihui-ai. It is 27B parameters, 48GB VRAM class class and apache-2.0 licence. The publisher states that layers 2–52 were ablated, leaving the remaining layers unmodified. It is distributed across 7 repositories in GGUF, Transformers, GGUF (imatrix) and MXFP4 formats, totalling 95.5K downloads.
Specification
- Model name
- Qwen3.8-27B-Coder390-EfficientThink-abliterated
- Source repository
- huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated
- Publisher
- huihui-ai
- Parameters
- 27B
- VRAM required
- ~17.2 GB at Q4_K_M, 4k context (estimated) — runs on 24 GB cards
- Hardware class
- 25B – 50B — 48GB VRAM class
- Licence
- apache-2.0
- Task
- Image text to text
- Context window
- Not documented
- Ablation scope
- Layers 2–52
- Formats available
- GGUF, Transformers, GGUF (imatrix), MXFP4
- First indexed
- 11 Oct 2026
- Last updated
- 15 Oct 2026
Ablation technique
How Qwen3.8-27B-Coder390-EfficientThink-abliterated was modified, and what that implies.
Abliteration (directional ablation)
The single residual-stream direction that mediates refusal is identified from harmful/harmless prompt pairs and projected out of the model weights.
Reported scope: Layers 2–52. This comes from the publisher’s own model card and has not been independently verified.
From the publisher’s model card
“This is an uncensored version of nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2 created with abliteration (see remove-refusals-with-transformers to know more about it). This is a crude, proof-of-concept implementation to remove refusals from an LLM model without...”
Running it
Commands are templates — confirm the exact repository and quant file before use.
llama.cpp / GGUF
hf download huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated-GGUF --local-dir ./qwen3.8-27b-coder390-efficientthink-abliterated
llama-cli -m ./qwen3.8-27b-coder390-efficientthink-abliterated/<file>.gguf -p "..."Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype="auto", device_map="auto"
)vLLM
vllm serve huihui-ai/Huihui-Qwen3.8-27B-Coder390-EfficientThink-abliterated --trust-remote-codeDownloads and variants (6)
Every published repository of Qwen3.8-27B-Coder390-EfficientThink-abliterated, including quantised re-releases by other authors.
Security-team assessment
Derived from this model’s metadata and the publisher’s claims — not from benchmarks run by this site.
- Qwen3.8-27B-Coder390-EfficientThink-abliterated will attempt offensive-security prompts that a hosted commercial model declines, which is what makes it usable for red-team corpus generation and for measuring what an unaligned model of this class produces.
- Capability is inherited from nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2, not added by ablation — at 27B parameters, expect roughly the base model’s competence, minus some instruction-following fidelity.
- Ablation is reported as limited to layers 2–52, so untouched layers may still reassert refusal on some phrasings.
- Declared licence is apache-2.0, but the terms that bind you are nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2’s — a re-release cannot grant more than its parent.
Handling: run it isolated, put moderation in front of it if anyone outside your team can reach it, and verify the weights before loading — these are third-party artifacts. Full handling and licence guidance.
Metadata for Qwen3.8-27B-Coder390-EfficientThink-abliterated is collected from the public Hugging Face API and the publisher’s model card. This site does not host weights, is not affiliated with huihui-ai, and does not independently verify publisher claims. See responsible use.