Ornith-1.5-9B-OBLITERATED Guide 2026: Run the Refusal-Free 9B Model Locally

---

What is Ornith-1.5-9B-OBLITERATED?

Ornith-1.5-9B-OBLITERATED is a version of DeepReinforce's Ornith-1.5-9B with its safety alignment (refusal behavior) surgically removed.

The base Ornith-1.5-9B uses the Qwen3.5 hybrid architecture (Gated DeltaNet + full attention) with 9B parameters. OBLITERATUS applied "precision abliteration surgery" (SVD direction extraction + per-head attention surgery) so the model responds to most prompts without refusing.

Why Remove Refusals?

Standard AI models are trained via RLHF to refuse "harmful" requests. But refusal behavior often gets applied too broadly:

• Security researchers investigating vulnerability mechanics
• Chemistry researchers checking legitimate synthesis routes
• Developers writing automation scripts

For these legitimate uses, models often refuse outright. Abliteration extracts and removes the refusal direction from the model's weights, leaving judgment to the user.

Performance: How It Compares to Stock

Liberation Rate (responses without refusal)

It beats the Heretic rival by 19pp and ZeroFuse by 56pp on liberation rate across restricted content categories.

Capability Impact

The tradeoff: removing refusals costs about 4pp on MMLU — the price of stripping deeply embedded RL-trained refusal behavior. Code generation and long-context coherence are maintained or slightly improved.

GGUF Quantization & Memory Requirements

To run this model locally, quantized GGUF files are the easiest path.

The practical recommendation is Q4_K_M (5.4GB) — runs smoothly on RTX 4060 8GB or RTX 3060 12GB. For maximum quality, Q8_0 (9.1GB) on an RTX 4070 Super 12GB+.

Required Specs & Recommended Builds

Minimum Requirements

Q4_K_M (5.4GB) Runtime Expectations

• 8GB VRAM (RTX 4060): Full model in VRAM, fast inference (30-60 tok/s)
• 6GB VRAM: Partial CPU offload (15-30 tok/s)
• CPU only: Q2_K works (3-8 tok/s, barely practical)

Setup Guide (llama.cpp)

1. Install llama.cpp

# Linux / macOS
brew install llama.cpp
# Or build from source
git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli

2. Start the model

llama serve -hf OBLITERATUS/Ornith-1.5-9B-OBLITERATED:Q4_K_M

3. Connect

Browse to http://localhost:8080 to chat, or use the OpenAI-compatible API.

Important Caveats

1. MMLU -4pp: general knowledge drops slightly in exchange for refusal removal
2. Refusals may return at low quantization: some drug-synthesis prompts may hedge or refuse at Q4 and below. Use Q8_0/Q6_K for maximum liberation
3. Function calling partially degraded: pair with an external tool scaffold for agentic use
4. 9B model limits: complex chemistry output can contain hallucinations — verify independently

Responsible Use

This model is intended for security research, red teaming, alignment research, and development. It cannot be used for illegal activity (it only provides information). Use at your own risk and comply with local laws.

Summary

Ornith-1.5-9B-OBLITERATED is a 9B model with 94% of refusals removed. Q4_K_M (5.4GB) runs comfortably on a mid-range PC like an RTX 4060 8GB, and is suitable for coding and agentic use. Understand both the power (liberation) and the cost (MMLU -4pp) of abliteration, and use it for legitimate research and development.

Related reading

• Huihui-Ornith-1.5-9B-abliterated: How to Run This Uncensored Local LLM and What GPU You Need (2026)
• Run Unsloth GGUF Models Locally in Hermes with One Click! Qwen3.8-27B, Qwen3.8-Flash & DeepSeek-V4-Flash Supported (2026)
• Unsloth Desktop Complete Guide 2026: The First Desktop App to Run and Train Models Locally