Vortex5/Pantheon-Reasoning-26B-A4B-1.1-heretic

VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Vortex5/Pantheon-Reasoning-26B-A4B-1.1-heretic is a 26 billion parameter decensored variant of Gryphe's Pantheon-Reasoning-26B-A4B-1.1, built on the Gemma 4 MoE architecture with a 32768 token context length. This model is specifically fine-tuned for enhanced roleplay quality by integrating extensive reasoning traces, allowing it to plan character responses and narrative beats. It excels at generating nuanced, prose-forward interactive fiction and general roleplay scenarios by simulating a writer's planning process.

Loading preview...

Overview

Vortex5/Pantheon-Reasoning-26B-A4B-1.1-heretic is a 26 billion parameter model derived from Gryphe's Pantheon-Reasoning-26B-A4B-1.1, utilizing the Gemma 4 MoE architecture. This version has been decensored using the Heretic tool with the Arbitrary-Rank Ablation (ARA) method, significantly reducing refusals from 100/100 to 13/100 compared to the original model, while maintaining a low KL divergence of 0.0455.

Key Capabilities

  • Enhanced Roleplay: Fine-tuned with a focus on character work, narrative planning, and tone consideration, aiming to improve roleplay quality.
  • Reasoning Integration: Incorporates extensive reasoning traces generated by DeepSeek 3.2, simulating a writer's planning process before generating responses.
  • Diverse Training Data: Trained on a mix of Pantheon roleplay data, general roleplay transcripts, text adventure content, and Claude Opus 4.6 reasoning traces for broad instruction-following and STEM capabilities.
  • Prose-Forward Style: Benefits from text adventure data to develop a more grounded, prose-forward writing style.

Good For

  • Interactive Fiction & Roleplay: Ideal for applications requiring models to engage in complex, character-driven roleplay and interactive storytelling.
  • Creative Writing Assistance: Useful for scenarios where a model needs to plan and generate nuanced narrative beats and character responses.
  • Decensored Applications: Suitable for use cases where a less restrictive response generation is desired, as indicated by its significantly reduced refusal rate.