andrevp/Ornith-1.0-9B-Heretic-Uncensored

Ornith-1.0-9B-Heretic-Uncensored: Abliterated Multimodal LLM

This model is an abliterated (uncensored) version of deepreinforce-ai/Ornith-1.0-9B, developed by andrevp. It leverages a Qwen 3.5-style hybrid architecture (24 linear-attention + 8 full-attention layers) with approximately 9 billion parameters and a 32768 token context length. The primary differentiator is the removal of the base model's refusal direction through directional ablation (weight orthogonalization), ensuring it no longer refuses requests without traditional retraining or fine-tuning that might degrade quality.

Key Capabilities & Features

• Uncensored Output: Designed to provide unfiltered responses by removing the refusal direction, making it suitable for research into model safety and behavior.
• Multimodal: Supports both vision and text inputs, indicating capabilities beyond pure text generation.
• Reasoning Model: Incorporates a <think>...</think> block at the start of assistant turns, suggesting enhanced internal reasoning processes.
• Directional Ablation: Utilizes a specific method based on Arditi et al. (2024) and Maxime Labonne's work to remove refusal tendencies by orthogonalizing weights in o_proj, out_proj, and down_proj components.
• Full-Precision: Provided in bf16 transformers/safetensors format.

Usage Considerations

This model is explicitly uncensored and carries significant warnings regarding its use. It may produce sensitive or controversial outputs and is not recommended for public-facing or high-security applications without careful monitoring. It is best suited for research and experimental use where unfiltered model behavior is desired, and users are responsible for ensuring legal and ethical compliance.