bowmanslayer/Ornith-1.5-9B-Uncensored

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 27, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

bowmanslayer/Ornith-1.5-9B-Uncensored is a 9 billion parameter language model based on the Ornith-1.5-9B architecture, which is a post-trained descendant of Qwen3.5-9B. This model has undergone directional ablation to remove safety alignment, allowing it to comply with requests that the original model would refuse. It maintains 86.86% of the base model's capability across 11 benchmarks, with a 32768 token context length, making it suitable for research and local inference where uncensored outputs are required.

Loading preview...

Ornith-1.5-9B-Uncensored: Uncensored Language Model

This model, developed by bowmanslayer, is a 9 billion parameter variant of the ornith-ai/Ornith-1.5-9B base model, which itself is a descendant of Qwen3.5-9B. Its primary distinction is the removal of safety alignment through a directional ablation technique, allowing it to generate content that the original model would refuse.

Key Capabilities & Features

  • Uncensored Output: Achieves a 0/23 refusal rate on held-out harmful prompts, compared to 20/23 for the original ablation method.
  • Capability Preservation: Retains 86.86% of the base model's performance across 11 benchmarks, with only a -0.59pp mean capability loss. Notably, TruthfulQA and BBH scores improved.
  • Ablation Method: Utilizes single-direction weight-space ablation (Arditi et al. 2024) on 64 residual-writing tensors, specifically excluding embed_tokens for effective refusal removal without significant capability degradation.
  • Vision Tower Untouched: The model's vision capabilities are inherited directly from Ornith-1.5-9B base, meaning multi-modal safety re-alignment has not been applied.
  • 32K Context Length: Supports a substantial context window of 32768 tokens.

Use Cases & Limitations

This model is intended for local inference and research by users who understand and accept the consequences of its uncensored nature. It will produce content that may be offensive, dangerous, or illegal. It is not recommended for public-facing deployment without robust external safety layers.

Known limitations include:

  • A small capability cost (mean -0.59pp), with larger drops in MATH-500 (-3.15) and HumanEval (-2.51).
  • Refusal removal is not exhaustively proven beyond the 23 tested prompts (English, single-turn).
  • The vision tower remains unaligned for safety.
  • The refusal metric is based on a small sample size (n=23).

Finetunes
3 models