Vortex5/G4-Moonlight-Dusk-26B-A4B-heretic

VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Vortex5/G4-Moonlight-Dusk-26B-A4B-heretic is a 26 billion parameter decensored language model, derived from Vortex5/G4-Moonlight-Dusk-26B-A4B using the Heretic v1.2.0 tool with Arbitrary-Rank Ablation (ARA) method. This model significantly reduces refusals compared to its original counterpart, making it optimized for creative writing tasks such as storytelling, roleplay, and atmospheric fiction. It maintains a context length of 32768 tokens, focusing on generating long-form narratives and emotion-forward interactions.

Loading preview...

Model Overview

Vortex5/G4-Moonlight-Dusk-26B-A4B-heretic is a 26 billion parameter language model created by Vortex5. It is a decensored version of the original G4-Moonlight-Dusk-26B-A4B model, achieved through the application of the Heretic v1.2.0 tool, specifically utilizing the Arbitrary-Rank Ablation (ARA) method with row-norm preservation.

Key Differentiators

This 'heretic' variant stands out due to its significant reduction in refusal rates compared to the base model. While the original model had 100/100 refusals in testing, this decensored version shows only 14/100 refusals, indicating a much more permissive output. The ablation process involved specific parameters, including start_layer_index (3), end_layer_index (22), preserve_good_behavior_weight (0.6457), steer_bad_behavior_weight (0.0003), overcorrect_relative_weight (1.0934), and neighbor_count (15).

Intended Use Cases

This model is specifically designed and optimized for creative text generation. Its primary intended uses include:

  • Storytelling: Generating structured long-form narratives.
  • Roleplay: Facilitating emotion-forward interactions.
  • Creative Writing: Producing atmospheric fiction.

Model Origin

The base G4-Moonlight-Dusk-26B-A4B model was created through a merge of several other models, including Darkhn/Gemma-4-26B-A4B-Animus-V14.1-FFT, zerofata/G4-MeroMero-26B-A4B, Locutusque/Esmeralda-Gemma4-26B-A4B, and Gryphe/Pantheon-Reasoning-26B-A4B-1.1, using a custom 'nexus' merge method.

Finetunes
1 models