alexokita/Phoenix-X-26B-A4B-Heretic

VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Phoenix-X-26B-A4B-Heretic is a 26 billion parameter language model created by alexokita, derived from a directional ablation merge of Vortex5/Phoenix-X-26B-A4B. This model significantly reduces refusals, achieving 21/100 compared to a baseline of 97/100, indicating a specialized focus on generating responses rather than declining prompts. With a 32768 token context length, it is optimized for tasks requiring reduced refusal rates and consistent output generation.

Loading preview...

Model Overview

Phoenix-X-26B-A4B-Heretic is a 26 billion parameter language model developed by alexokita, based on a directional ablation merge of the Vortex5/Phoenix-X-26B-A4B model. This model was created using the Heretic merging technique, specifically designed to modify model behavior through targeted ablation.

Key Characteristics

  • Reduced Refusal Rate: A primary differentiator of this model is its significantly lower refusal rate, recorded at 21/100 trials compared to a baseline of 97/100. This indicates a strong bias towards generating responses rather than declining prompts.
  • KL Divergence: The merge process resulted in a KL divergence of 0.0225, suggesting a controlled and specific modification from its base model.
  • Context Length: The model supports a substantial context length of 32768 tokens, enabling it to process and generate longer sequences of text.
  • GGUF Quantization: Optimized GGUF i1 quants are available under alexokita/Phoenix-X-26B-A4B-Heretic-i1-GGUF for efficient deployment.

Use Cases

This model is particularly well-suited for applications where minimizing model refusals is critical, such as:

  • Interactive Agents: Scenarios requiring consistent and non-refusal responses from an AI.
  • Content Generation: Tasks where the model needs to always attempt to generate content, even for challenging prompts.
  • Specialized Chatbots: Chatbots designed to maintain engagement and avoid 'I cannot answer that' type responses.