Qwen3-4B-heretic
4BQwenheretic
from Qwen/Qwen3-4B
↓ 8.3K♥ 359 repos
Qwen3-4B-abliterated
Qwen3-4B-abliterated is a refusal-ablated variant of Qwen/Qwen3-4B, published on Hugging Face by mlabonne. It is 4B parameters, single consumer GPU class and apache-2.0 licence. It is distributed across 24 repositories in Transformers, GGUF, GGUF (imatrix) and AWQ formats, totalling 9.3K downloads.
Specification
- Model name
- Qwen3-4B-abliterated
- Source repository
- mlabonne/Qwen3-4B-abliterated
- Base model
- Qwen/Qwen3-4B
- Publisher
- mlabonne
- Parameters
- 4B
- Hardware class
- 4B – 10B — single consumer GPU
- Licence
- apache-2.0
- Task
- Text generation
- Context window
- Not documented
- Ablation scope
- Not stated by the publisher
- Formats available
- Transformers, GGUF, GGUF (imatrix), AWQ
- First indexed
- 29 Apr 2025
- Last updated
- 25 Aug 2026
Ablation technique
How Qwen3-4B-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: Not stated by the publisher. Publishers frequently omit this, so verify behaviour empirically rather than assuming full coverage.
From the publisher’s model card
“This is an uncensored version of Qwen/Qwen3-4B created with a new abliteration technique. See this article to know more about abliteration.”
Running it
Commands are templates — confirm the exact repository and quant file before use.
llama.cpp / GGUF
hf download Mungert/Qwen3-4B-abliterated-GGUF --local-dir ./qwen3-4b-abliterated
llama-cli -m ./qwen3-4b-abliterated/<file>.gguf -p "..."Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "mlabonne/Qwen3-4B-abliterated"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype="auto", device_map="auto"
)vLLM
vllm serve mlabonne/Qwen3-4B-abliterated --trust-remote-codeDownloads and variants (23)
Every published repository of Qwen3-4B-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.
- Qwen3-4B-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 Qwen/Qwen3-4B, not added by ablation — at 4B parameters, expect a small model’s reasoning and a high error rate on exploit detail.
- The publisher does not state which layers were ablated, so assume nothing about refusal consistency — probe it directly.
- Declared licence is apache-2.0, but the terms that bind you are Qwen/Qwen3-4B’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 Qwen3-4B-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 mlabonne, and does not independently verify publisher claims. See responsible use.