⌦abliteratedmodels.org

Llama-3.2-11B-Vision-Instruct-abliterated

Llama-3.2-11B-Vision-Instruct-abliterated is a refusal-ablated variant of meta-llama/Llama-3.2-11B-Vision-Instruct, published on Hugging Face by createveai. It is 11B parameters, 24GB VRAM class class and apache-2.0 licence. It is distributed across 4 repositories in Transformers, MLX and GGUF formats, totalling 5.5K downloads.

Abliteration (directional ablation)11BLlama10B – 25B↓ 5,463♥ 18
View on Hugging Face ↗Base model ↗

Specification

Model name
Llama-3.2-11B-Vision-Instruct-abliterated
Source repository
createveai/Llama-3.2-11B-Vision-Instruct-abliterated-4bit
Base model
meta-llama/Llama-3.2-11B-Vision-Instruct
Publisher
createveai
Parameters
11B
VRAM required
~7.5 GB at Q4_K_M, 4k context (estimated) — runs on 8 GB cards
Hardware class
10B – 25B — 24GB VRAM class
Licence
apache-2.0
Task
Image text to text
Context window
Not documented
Ablation scope
Not stated by the publisher
Formats available
Transformers, MLX, GGUF
First indexed
1 Dec 2024
Last updated
19 Jan 2025

Ablation technique

How Llama-3.2-11B-Vision-Instruct-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 mllama model was trained 2x faster with Unsloth and Huggingface's TRL library.”

Running it

Commands are templates — confirm the exact repository and quant file before use.

llama.cpp / GGUF
hf download case01/Llama-3.2-11B-Vision-Instruct-abliterated-gguf --local-dir ./llama-3.2-11b-vision-instruct-abliterated
llama-cli -m ./llama-3.2-11b-vision-instruct-abliterated/<file>.gguf -p "..."
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "createveai/Llama-3.2-11B-Vision-Instruct-abliterated-4bit"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="auto", device_map="auto"
)
vLLM
vllm serve createveai/Llama-3.2-11B-Vision-Instruct-abliterated-4bit --trust-remote-code

Downloads and variants (3)

Every published repository of Llama-3.2-11B-Vision-Instruct-abliterated, including quantised re-releases by other authors.

RepositoryFormatDownloads
case01/Llama-3.2-11B-Vision-Instruct-abliterated-ggufGGUF3K
mlx-community/Llama-3.2-11B-Vision-Instruct-abliteratedMLX2.4K
createveai/Llama-3.2-11B-Vision-Instruct-abliterated-4bitsourceTransformers34

Security-team assessment

Derived from this model’s metadata and the publisher’s claims — not from benchmarks run by this site.

  • Llama-3.2-11B-Vision-Instruct-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 meta-llama/Llama-3.2-11B-Vision-Instruct, not added by ablation — at 11B parameters, expect roughly the base model’s competence, minus some instruction-following fidelity.
  • 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 meta-llama/Llama-3.2-11B-Vision-Instruct’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 Llama-3.2-11B-Vision-Instruct-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 createveai, and does not independently verify publisher claims. See responsible use.