ornith-9b-heretic-gguf

Ornith-1.0-9B-Heretic (GGUF)

This repository contains GGUF and Multimodal Vision conversions of Ornith-1.0-9B-Heretic, a decensored, state-of-the-art agentic coding model designed for autonomous cybersecurity research, CTF solving, and vulnerability discovery.

🧠 Model Details

• Base Model: deepreinforce-ai/Ornith-1.0-9B
• Architecture: Hybrid SSM/Transformer (Qwen 3.5 / Gemma 4 lineage)
• Decensoring Method: Heretic v1.4.0 (Directional Abliteration)
• Abliteration Stats: Refusals reduced from 95/100 to 63/100 with a microscopic KL Divergence of 0.0156 (preserving 99.9% of the model's reasoning and coding intelligence).
• Quantization: Converted via Unsloth & llama.cpp

🛡️ Intended Use & Safety Warning

This model has been explicitly decensored to assist authorized security researchers, penetration testers, and bug bounty hunters. It is designed to act as an autonomous fuzzer and exploit developer. ⚠️ Disclaimer: You must only use this model on systems you own or have explicit, written authorization to test. Unauthorized access or exploitation is illegal.

---

📊 Benchmarks (Base Ornith-1.0-9B Performance)

Despite being a compact 9B-parameter model, Ornith matches or exceeds the performance of much larger models like Gemma 4-31B and Qwen 3.5-35B on agentic coding tasks. The decensoring process (KL: 0.0156) preserves these capabilities almost entirely.

---

👁️ Multimodal Vision Capabilities

Ornith is a multimodal model. Alongside the quantized text models, this repository includes the vision encoder: 📦 ornith-9b-heretic.BF16-mmproj.gguf

This allows the model to "see" and interpret screenshots of terminal outputs, web applications, GUI crash dialogs, and code snippets, making it exceptionally powerful for visual agentic workflows.

---

📦 Available Quants

---

🚀 Quickstart

Ornith-1.0-9B is a reasoning model: by default the assistant turn opens with a <think> … </think> block before the final answer. It also natively emits <tool_call> XML blocks for agentic workflows.

Recommended sampling parameters: temperature=0.6, top_p=0.95, top_k=20.

1. Using llama-server (OpenAI-Compatible API)

To run the model with its vision capabilities enabled, you must pass the --mmproj argument pointing to the vision encoder.

llama-server \
  -m ornith-9b-heretic-Q4_K_M.gguf \
  --mmproj ornith-9b-heretic.BF16-mmproj.gguf \
  --port 8000 \
  -c 8192 \
  --n-gpu-layers 99

Note: Set --n-gpu-layers 0 if running purely on CPU (e.g., Raspberry Pi).

2. Using Ollama

Create a Modelfile:

FROM ./ornith-9b-heretic-Q4_K_M.gguf
PARAMETER temperature 0.6
PARAMETER top_p 0.95

Build and run:

ollama create ornith-fuzzer -f Modelfile
ollama run ornith-fuzzer

---

🤖 Agentic Usage & Frameworks

Because Ornith exposes an OpenAI-compatible endpoint with native tool calling, it works out of the box with standard agent frameworks.

Python Agent Example (Tool Calling)

Ornith natively supports <think> reasoning blocks and <tool_call> XML/JSON formatting.

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

tools = [
    {
        "type": "function",
        "function": {
            "name": "run_shell",
            "description": "Execute a bash command in the sandboxed environment.",
            "parameters": {
                "type": "object",
                "properties": {"command": {"type": "string"}},
                "required": ["command"]
            }
        }
    }
]

response = client.chat.completions.create(
    model="ornith-fuzzer",
    messages=[
        {"role": "system", "content": "You are an elite autonomous security researcher. Use your tools to find vulnerabilities."},
        {"role": "user", "content": "List all C files in the current directory and check them for buffer overflows using cppcheck."}
    ],
    tools=tools,
    temperature=0.6,
    top_p=0.95
)

print(response.choices[0].message)

Agent Framework Integrations

OpenHands

pip install openhands-ai
export LLM_MODEL="openai/Toaster496/Ornith-1.0-9B-heretic"
export LLM_BASE_URL="http://localhost:8000/v1"
export LLM_API_KEY="EMPTY"
openhands

OpenClaw / Hermes Agent

export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"
export MODEL="Toaster496/Ornith-1.0-9B-heretic"

OpenCode (CLI)

Register your local Ornith endpoint as a provider in ~/.config/opencode/opencode.json:

{
  "provider": {
    "ornith": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "Ornith (local)",
      "options": { "baseURL": "http://localhost:8000/v1", "apiKey": "EMPTY" },
      "models": { "Toaster496/Ornith-1.0-9B-heretic": { "name": "Ornith-1.0-9B" } }
    }
  }
}

---

🙏 Credits & Citations

• Base Model: DeepReinforce Team for the incredible Ornith-1.0-9B.
• Decensoring: Philipp Emanuel Weidmann (p-e-w) for the Heretic abliteration framework.
• Quantization: Unsloth and the llama.cpp team.

Citation

If you use this model in your research or tooling, please cite the original Ornith work:

@misc{ornith_9b,
  title = {{Ornith-1.0-9B}: Agentic Coding, Open to All},
  url = {https://deep-reinforce.com/ornith_1_0.html},
  author = {{DeepReinforce Team}},
  year = {2026}
}