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

Hugging Face
VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 27, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Warm

ansulev/Ornith-1.0-9B-Heretic-Uncensored is a 9 billion parameter Qwen 3.5-style hybrid multimodal vision and text model, based on deepreinforce-ai/Ornith-1.0-9B. This model has undergone directional ablation to remove its refusal direction, making it an uncensored version. It retains the base model's reasoning capabilities, opening assistant turns with a block, and is suitable for use cases requiring unfiltered outputs.

Loading preview...

Ornith-1.0-9B-Heretic-Uncensored Overview

This model is an abliterated (uncensored) version of deepreinforce-ai/Ornith-1.0-9B, developed by ansulev. It utilizes a Qwen 3.5-style hybrid architecture, featuring 32 layers (24 linear-attention and 8 full-attention layers) and multimodal vision + text capabilities. With approximately 9 billion parameters, it maintains the reasoning structure of the base model, where assistant responses begin with a <think>...</think> block.

Key Differentiator: Abliteration

The primary distinction of this model is the removal of its refusal direction through directional ablation. This process involved:

  • Data Collection: Analyzing residual-stream activations from harmful and harmless prompts.
  • Refusal Direction Identification: Computing mean differences between harmful and harmless activations to identify a single best refusal direction (layer 28).
  • Weight Orthogonalization: Applying this direction to the o_proj/out_proj and down_proj weight matrices across all layers, permanently preventing the model from generating refusal-based outputs. This method is based on Arditi et al. (2024) and Maxime Labonne's abliteration guide.

Usage Warnings

As an uncensored model, users should be aware of:

  • Reduced Safety Filtering: Increased risk of sensitive or controversial outputs.
  • Suitability: Not recommended for public-facing or high-security applications without careful monitoring.
  • Responsibility: Users are solely responsible for legal and ethical compliance.

This model is best suited for research and experimental use where unfiltered outputs are required, and the user accepts the associated risks.

Finetunes
1 models