Naphula/Goetia-26B-A4B-v1.4

VISIONConcurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Naphula/Goetia-26B-A4B-v1.4 is a 26 billion parameter language model based on the Gemma-4 architecture, created by merging multiple pre-trained models using the MoE DELLA method. This model integrates various specialized Gemma-4-26B-A4B variants, including those focused on reasoning and style-tuning, to enhance overall performance. With a 32768 token context length, it is designed for complex tasks requiring a blend of diverse capabilities derived from its constituent models.

Loading preview...

Goetia 26B A4B v1.4 Overview

Goetia 26B A4B v1.4 is a 26 billion parameter language model built upon the Gemma-4 architecture. It was developed by Naphula through a sophisticated merging process using the MoE DELLA method, which combines the strengths of several specialized Gemma-4-26B-A4B models. The base model for this merge was google/gemma-4-26B-A4B.

Key Capabilities & Features

  • Advanced Merging Technique: Utilizes the MoE DELLA merge method, allowing for a blend of expert models to enhance overall performance.
  • Diverse Model Integration: Incorporates contributions from models like BeaverAI/Orion-26B-A4B-v1b-GGUF, Darkhn/Gemma-4-26B-A4B-Animus-V14.1-FFT, Gryphe/Pantheon-Reasoning-26B-A4B-1.1, and Gryphe/Gemma-4-26B-A4B-StyleTune-V2, suggesting a focus on varied capabilities including reasoning and stylistic generation.
  • Gemma-4 Architecture: Benefits from the underlying strengths of the Gemma-4 base model.
  • Configurable Merging: The merge configuration details, including specific weight filters and density parameters, are provided for transparency and reproducibility.

Important Considerations

  • Censorship: This model is not uncensored and exhibits standard refusal behaviors and jailbreak resistance. However, it can be decensored using the Heretic tool, with a provided reproduce.json example for ablation using ARA.
  • Patches Applied: The v1.4 iteration includes critical patches for auto.py, plan.py, and gemma4.json to ensure correct merging, particularly for lm_head and embed_tokens in Gemma 4 configurations.

Finetunes
1 models