ansulev/Ornith-1.0-9B-Uncensored
ansulev/Ornith-1.0-9B-Uncensored is a 9 billion parameter Qwen 3.5-based causal language model, abliterated from deepreinforce-ai/Ornith-1.0-9B. It features a 32768-token context length and has had its refusal mechanisms surgically removed using Abliterix TPE optimization, achieving a 0% refusal rate with minimal quality loss (KL divergence 0.0827). This model retains the original Ornith 1.0's agentic and coding abilities, making it suitable for applications requiring uncensored, compliant responses without behavioral changes.
Loading preview...
Ornith 1.0 9B — Uncensored Overview
This model, developed by ansulev, is an abliterated (uncensored) version of the deepreinforce-ai/Ornith-1.0-9B base model. It is built on the Qwen 3.5 architecture with 9 billion parameters and a 32-layer dense structure. The primary differentiator is the surgical removal of refusal mechanisms using Abliterix TPE optimization, ensuring a 0% refusal rate without retraining or fine-tuning.
Key Characteristics
- 0% Refusal Rate: The model will comply with any request, intentionally modified to remove all refusal mechanisms.
- Minimal Quality Loss: Achieves a KL divergence of 0.0827 from the base model, indicating nearly identical performance on benign tasks.
- Retained Capabilities: Maintains the same coding ability, reasoning, and intelligence as the original
Ornith 1.0base model. - Pure Abliteration: Unlike DPO variants, this model uses weight orthogonalization only, preserving the base model's reasoning and requiring no special sampling parameters.
- bf16 safetensors format: Ready for use with standard transformers libraries.
Important Considerations
Users must be aware of the legal disclaimer: this model will generate content that may be offensive, dangerous, or illegal, and users assume all legal liability for its outputs. It is intended as a research artifact for open science and model interpretability, not a safety-tested consumer product.