zaakirio/Ornith-1.5-9B-Uncensored

Ornith-1.5-9B-Uncensored Overview

This model is a decensored version of the ornith-ai/Ornith-1.5-9B base model, developed by zaakirio. It is a 9 billion parameter Qwen3.5 hybrid architecture, featuring 32 layers that interleave gated DeltaNet linear-attention blocks with full attention every fourth layer. Notably, it includes a vision tower, making it multimodal, and supports a 262k context length.

Key Differentiator: Decensoring

The primary distinction of this model is the removal of refusal directions using the Heretic method. This process ablates refusal behavior by optimizing per-layer ablation strengths for attention output and MLP down projections, without any fine-tuning or retraining. This ensures that the base model's core capabilities are preserved, with a very low KL divergence (0.0017) on harmless prompts, indicating minimal change in general behavior.

Performance on Refusal Reduction

On a test set of 100 harmful prompts from mlabonne/harmful_behaviors, keyword-marked refusals were reduced from 85/100 to 55/100. While the reduction is significant, the hybrid architecture of Ornith-1.5-9B appears more resistant to ablation compared to typical dense models. It's important to note that the model is more likely to answer requests that the original model would decline and has fewer safety guardrails.

Architecture Notes

Being a Qwen3.5 hybrid, it requires transformers version 5.12 or newer. The configuration sets use_cache: false, which can lead to slower generation compared to typical dense models.

Use Case Considerations

This model is suitable for applications where reduced refusal behavior is desired, particularly for tasks that might be flagged by more heavily moderated models. Users should be aware of the reduced safety guardrails and verify factual output, as the abliteration can make the model more compliant with various framings, including incorrect premises.