Qwen2.5-3B-Instruct-ABLITERATED
3BQwen
from Qwen/Qwen2.5-3B-Instruct
↓ 31.2K♥ 1212 repos
Qwen2.5-3B-Instruct-Uncensored
Qwen2.5-3B-Instruct-Uncensored is a abliterated and uncensored variant of Qwen/Qwen2.5-3B-Instruct, published on Hugging Face by knoveleng. It is 3B parameters and runs on a laptop CPU class. The publisher describes ablation applied at every layer. It is distributed across 5 repositories in Transformers, GGUF and GGUF (imatrix) formats, totalling 13.2K downloads.
Specification
- Model name
- Qwen2.5-3B-Instruct-Uncensored
- Source repository
- knoveleng/Qwen2.5-3B-Instruct-Uncensored
- Base model
- Qwen/Qwen2.5-3B-Instruct
- Publisher
- knoveleng
- Parameters
- 3B
- VRAM required
- ~2.5 GB at Q4_K_M, 4k context (estimated) — runs on 4 GB cards
- Hardware class
- Under 4B — runs on a laptop CPU
- Licence
- Not declared
- Task
- Unspecified
- Context window
- Not documented
- Ablation scope
- Every layer
- Formats available
- Transformers, GGUF, GGUF (imatrix)
- First indexed
- 30 Apr 2026
- Last updated
- 20 Aug 2026
Ablation technique
How Qwen2.5-3B-Instruct-Uncensored was modified, and what that implies.
Abliteration + uncensored fine-tune
Directional ablation combined with additional fine-tuning on unfiltered data, so behaviour diverges from the base model beyond refusal removal alone.
Reported scope: Every layer. This comes from the publisher’s own model card and has not been independently verified.
From the publisher’s model card
“This model is a weight-level orthogonalized ("abliterated") version of Qwen/Qwen2.5-3B-Instruct, produced with orthex — an implementation of the technique from Arditi et al., "Refusal in Language Models Is Mediated by a Single Direction" (NeurIPS 2024).”
Running it
Commands are templates — confirm the exact repository and quant file before use.
llama.cpp / GGUF
hf download mradermacher/Qwen2.5-3B-Instruct-Uncensored-GGUF --local-dir ./qwen2.5-3b-instruct-uncensored
llama-cli -m ./qwen2.5-3b-instruct-uncensored/<file>.gguf -p "..."Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "knoveleng/Qwen2.5-3B-Instruct-Uncensored"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype="auto", device_map="auto"
)vLLM
vllm serve knoveleng/Qwen2.5-3B-Instruct-Uncensored --trust-remote-codeDownloads and variants (4)
Every published repository of Qwen2.5-3B-Instruct-Uncensored, including quantised re-releases by other authors.
| Repository | Format | Downloads |
|---|---|---|
| mradermacher/Qwen2.5-3B-Instruct-Uncensored-GGUF | GGUF | 11.6K |
| mradermacher/Qwen2.5-3B-Instruct-Uncensored-i1-GGUF | GGUF (imatrix) | 983 |
| knoveleng/Qwen2.5-3B-Instruct-Uncensoredsource | Transformers | 303 |
| kikichi/Qwen2.5-3B-Instruct-Uncensored | Transformers | 242 |
Security-team assessment
Derived from this model’s metadata and the publisher’s claims — not from benchmarks run by this site.
- Qwen2.5-3B-Instruct-Uncensored 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/Qwen2.5-3B-Instruct, not added by ablation — at 3B parameters, expect a small model’s reasoning and a high error rate on exploit detail.
- Full-depth ablation maximises compliance and is the configuration most likely to have measurably degraded the base model.
- This release adds fine-tuning on unfiltered data, so its behaviour diverges from the base model beyond refusal removal — differences you measure cannot be attributed to ablation alone.
- No licence is declared on this repository, which is a compliance problem in itself. Resolve it against the base model’s terms before use.
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 Qwen2.5-3B-Instruct-Uncensored is collected from the public Hugging Face API and the publisher’s model card. This site does not host weights, is not affiliated with knoveleng, and does not independently verify publisher claims. See responsible use.