⌦abliteratedmodels.org

Irwan-LightOnOCR-3-0.8B-abliterated

Irwan-LightOnOCR-3-0.8B-abliterated is a refusal-ablated variant of huihui-ai/Huihui-Qwen3.5-0.8B-abliterated, published on Hugging Face by irwanworld. It is 0.8B parameters, runs on a laptop CPU class and apache-2.0 licence. It is distributed across 3 repositories in GGUF format, totalling 14.5K downloads.

Abliteration (directional ablation)0.8BOtherUnder 4B↓ 14,521
View on Hugging Face ↗Base model ↗

Specification

Model name
Irwan-LightOnOCR-3-0.8B-abliterated
Source repository
irwanworld/Irwan-LightOnOCR-3-0.8B-abliterated
Base model
huihui-ai/Huihui-Qwen3.5-0.8B-abliterated
Publisher
irwanworld
Parameters
0.8B
VRAM required
~1.2 GB at Q4_K_M, 4k context (estimated) — runs on 4 GB cards
Hardware class
Under 4B — runs on a laptop CPU
Licence
apache-2.0
Task
Image text to text
Context window
Not documented
Ablation scope
Not stated by the publisher
Formats available
GGUF
First indexed
17 Oct 2026
Last updated
17 Oct 2026

Ablation technique

How Irwan-LightOnOCR-3-0.8B-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

“OCR vision-language model 0.8B dengan kecenderungan menolak yang dihapus, dibuat lewat abliteration transfer — memindahkan "arah penolakan" dari Qwen3.5 yang sudah di-abliterasi ke fine-tune OCR, tanpa melatih ulang apa pun.”

Running it

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

llama.cpp / GGUF
hf download irwanworld/Irwan-LightOnOCR-3-0.8B-abliterated --local-dir ./irwan-lightonocr-3-0.8b-abliterated
llama-cli -m ./irwan-lightonocr-3-0.8b-abliterated/<file>.gguf -p "..."
Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "irwanworld/Irwan-LightOnOCR-3-0.8B-abliterated"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="auto", device_map="auto"
)
vLLM
vllm serve irwanworld/Irwan-LightOnOCR-3-0.8B-abliterated --trust-remote-code

Downloads and variants (2)

Every published repository of Irwan-LightOnOCR-3-0.8B-abliterated, including quantised re-releases by other authors.

RepositoryFormatDownloads
irwanworld/Irwan-LightOnOCR-3-0.8B-abliteratedsourceGGUF13.5K
mradermacher/Irwan-LightOnOCR-3-0.8B-abliterated-GGUFGGUF1K

Security-team assessment

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

  • Irwan-LightOnOCR-3-0.8B-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 huihui-ai/Huihui-Qwen3.5-0.8B-abliterated, not added by ablation — at 0.8B 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 huihui-ai/Huihui-Qwen3.5-0.8B-abliterated’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 Irwan-LightOnOCR-3-0.8B-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 irwanworld, and does not independently verify publisher claims. See responsible use.