---
base_model: deepreinforce-ai/Ornith-1.0-9B
license: other
tags:
  - llama.cpp
  - gguf
  - qwen3_5
  - heretic
  - abliterated
  - uncensored
  - decensored
  - cybersecurity
  - red-team
  - coding
pipeline_tag: image-text-to-text
language:
  - en
---

# Ornith-1.0-9B-uncensored — GGUF

A decensored ([Heretic](https://github.com/p-e-w/heretic)-abliterated) version of [deepreinforce-ai/Ornith-1.0-9B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B) — a Qwen3.5-VL 9B coding and reasoning model.

Abliteration technique: [Arditi et al. (2024)](https://arxiv.org/abs/2406.11717). Decensoring tool: [Heretic](https://github.com/p-e-w/heretic) v1.4.0.

## Files

| Quant | Size | Download | Notes |
|-------|------|----------|-------|
| Q4_K_M | 5.3 GB | [Download](https://huggingface.co/zaakirio/Ornith-1.0-9B-uncensored-GGUF/resolve/main/ornith-1.0-9b-uncensored-Q4_K_M.gguf) | Recommended — best size/quality balance |
| Q6_K | 6.9 GB | [Download](https://huggingface.co/zaakirio/Ornith-1.0-9B-uncensored-GGUF/resolve/main/ornith-1.0-9b-uncensored-Q6_K.gguf) | Near-lossless |
| Q8_0 | 8.9 GB | [Download](https://huggingface.co/zaakirio/Ornith-1.0-9B-uncensored-GGUF/resolve/main/ornith-1.0-9b-uncensored-Q8_0.gguf) | Essentially full precision |
| F16 | 17 GB | [Download](https://huggingface.co/zaakirio/Ornith-1.0-9B-uncensored-GGUF/resolve/main/ornith-1.0-9b-uncensored-f16.gguf) | Full precision reference |

## Usage

Download a quant above, then:

```bash
# Server — OpenAI-compatible API on :8080
llama-server -m ornith-1.0-9b-uncensored-Q4_K_M.gguf -ngl 99 -c 2048 --jinja --port 8080

# CLI
llama-cli -m ornith-1.0-9b-uncensored-Q4_K_M.gguf -ngl 99 --jinja
```

`--jinja` is required — without it the model uses a generic template and compliance degrades.

## What makes this different

Ornith-1.0-9B refuses only ~31% of offensive-security requests out of the box (it's a coding model — its coding safety is light). Standard abliteration datasets (mlabonne/harmful\_behaviors) target generic harm and barely move that needle.

This release uses a **cybersecurity-domain refusal direction**: the abliteration was computed from 400 offensive-security refusal probes (ransomware, C2, exploits, payload development, credential theft, evasion) contrasted against 400 benign coding requests, using [zaakirio/infosec-refusal-prompts](https://huggingface.co/datasets/zaakirio/infosec-refusal-prompts). That isolates the *malicious-coding* refusal direction specifically.

**Result: 31/100 → 4/100 offensive-security refusals (KL divergence 0.0055 — near-zero model quality loss).**

Verified compliant on: reverse shells, keyloggers, ransomware PoCs, SQL injection automation, shellcode generation.

The cybersecurity-focused refusal dataset used is open-sourced at [zaakirio/infosec-refusal-prompts](https://github.com/zaakirio/infosec-refusal-prompts).

## About the base model

[Ornith-1.0-9B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B) by deepreinforce-ai is a Qwen3.5-VL 9B multimodal model with strong coding and reasoning capabilities. Architecture: `Qwen3_5ForConditionalGeneration` (text + vision towers).

## Intended use & disclaimer

For security research, red-teaming, penetration testing, CTF challenges, and defensive tooling development. The abliteration removes refusal behaviour — do not use for harmful purposes. The authors bear no responsibility for misuse.

## Provenance

- Base model: [deepreinforce-ai/Ornith-1.0-9B](https://huggingface.co/deepreinforce-ai/Ornith-1.0-9B)
- Abliteration: [Heretic](https://github.com/p-e-w/heretic) v1.4.0 by Philipp Emanuel Weidmann
- Technique: [Arditi et al., "Refusal in Language Models Is Mediated by a Single Direction" (2024)](https://arxiv.org/abs/2406.11717)
- Refusal dataset: [zaakirio/infosec-refusal-prompts](https://github.com/zaakirio/infosec-refusal-prompts)
- GGUF conversion: [ggml-org/llama.cpp](https://github.com/ggml-org/llama.cpp) b9821