Toaster496/Ornith-1.0-9B-heretic
Toaster496/Ornith-1.0-9B-heretic is a 9 billion parameter decensored variant of the deepreinforce-ai/Ornith-1.0-9B agentic coding model, maintaining 99.9% of its original reasoning and coding intelligence. This model has had its safety alignment and refusal vectors removed using the Heretic framework, specifically designed for cybersecurity research, penetration testing, CTF, and bug bounty hunting. It excels as an autonomous fuzzer and exploit developer without triggering standard safety refusals on technical cybersecurity prompts, while supporting a 32768 token context length.
Loading preview...
Ornith-1.0-9B-Heretic: Decensored Agentic Coding Model
Ornith-1.0-9B-Heretic is a specialized 9 billion parameter model derived from the deepreinforce-ai/Ornith-1.0-9B base model. Its primary distinction is the removal of safety alignment and refusal vectors using the Heretic directional abliteration framework. This process has preserved 99.9% of the original model's reasoning and coding intelligence (KL Divergence: 0.0156), while significantly reducing refusals on harmful behaviors from 95/100 to 63/100 on standard benchmarks, and maintaining near-perfect compliance on technical/coding prompts.
Key Capabilities
- Decensored Agentic Coding: Explicitly designed to bypass standard safety refusals for technical cybersecurity tasks.
- High Reasoning Fidelity: Retains the advanced agentic tool-calling and
<think>reasoning capabilities of its base model. - Cybersecurity Focus: Optimized for tasks such as autonomous fuzzing, exploit development, and penetration testing.
- Efficient Deployment: As the most lightweight member of the Ornith family, it's suitable for single-GPU deployment.
- Strong Coding Benchmarks: The base Ornith-1.0-9B model achieves competitive performance on agentic coding benchmarks like Terminal-Bench 2.1 (43.1), SWE-bench Verified (69.4), and SWE-bench Pro (42.9).
Good For
- Authorized Security Research: Ideal for security researchers, penetration testers, CTF players, and bug bounty hunters.
- Exploit Development: Generating exploits and fuzzer payloads without content restrictions.
- Technical Cybersecurity Prompts: Handling prompts related to system vulnerabilities, reverse engineering, and ethical hacking where other models might refuse.
⚠️ Safety Warning: This model is explicitly decensored. It must only be used on systems for which you have explicit, written authorization to test. Unauthorized use is illegal.