Naphula/Goetia-26B-A4B-v1.3-Absolute-Heretic-ARA

VISIONConcurrency Cost:2Model Size:26BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Naphula/Goetia-26B-A4B-v1.3-Absolute-Heretic-ARA is a 26 billion parameter language model based on the Gemma 4 architecture, created by Naphula. This model is a decensored variant of Naphula/Goetia-26B-A4B-v1.3, achieved through Arbitrary-Rank Ablation (ARA) using the Heretic v1.2.0 method. It is specifically optimized to produce narratives and roleplay content that may include violent and graphic erotic material, with a context length of 32768 tokens.

Loading preview...

Model Overview

Naphula/Goetia-26B-A4B-v1.3-Absolute-Heretic-ARA is a 26 billion parameter model built upon the Gemma 4 architecture, developed by Naphula. It is a merged model, combining google/gemma-4-26B-A4B with several other Gemma 4-based models using the MoE DELLA merge method. This particular version is a decensored iteration of the original Naphula/Goetia-26B-A4B-v1.3, achieved through the Arbitrary-Rank Ablation (ARA) method using Heretic v1.2.0.

Key Characteristics

  • Decensored Content Generation: Specifically engineered to bypass refusal mechanisms, allowing for the generation of violent and graphic erotic narratives and roleplay content.
  • High Reproducibility: The model's creation process is fully reproducible, with detailed guides and scripts provided, including specific environment setups and patches for Heretic.
  • Optimized Ablation: Utilizes a "Surgical Narrowing" strategy during the ARA process, focusing on specific layers and weight ratios to achieve decensoring while maintaining model coherence.
  • Blackwell Hardware Optimization: The ablation process was specifically tuned and executed on NVIDIA RTX 6000 Blackwell hardware (96GB VRAM) for lossless 16-bit precision.

Use Cases

  • Creative Writing & Roleplay: Ideal for applications requiring unrestricted and explicit content generation, particularly in genres involving violence or graphic eroticism.
  • Research into Model Alignment & Safety: Can be used by researchers to study the effects of decensoring techniques and the boundaries of LLM safety filters.