rpDungeon/Gemma4-31b-Gembrain-Equinox

VISIONConcurrency Cost:2Model Size:31BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 23, 2026License:gemmaArchitecture:Transformer0.0K Cold

rpDungeon/Gemma4-31b-Gembrain-Equinox is a 31 billion parameter Gemma-4 based model, created by rpDungeon, jaxxks, twisted, and toasty, specifically designed for creative writing tasks. This V2 Fisher-protected community merge combines two Gemma-4 31B creative-writing variants, Gembrain and Equinox, on top of Google's stock gemma-4-31b-it. It features a modified chat template to improve instruct-following and prompt-attention, making it suitable for nuanced creative text generation with a context length of 32768 tokens.

Loading preview...

Model Overview

rpDungeon/Gemma4-31b-Gembrain-Equinox is a 31 billion parameter model built on the Gemma-4 architecture, resulting from a V2 Fisher-protected community merge. This model integrates two distinct community-developed Gemma-4 31B creative-writing variants, Gembrain and Equinox, with Google's base gemma-4-31b-it model. The merge utilizes a TIES-style approach with Fisher importance and layer-importance damping to preserve instruct-following capabilities while blending the unique styles of the community variants.

Key Capabilities

  • Enhanced Creative Writing: Blends the prompt-attention reinforcement of Gembrain with the neutral/realistic style of Equinox for diverse creative outputs.
  • Improved Instruct-Following: Incorporates Fisher importance and layer-importance damping during the merge process to protect critical instruct-following parameters.
  • Optimized Chat Template: Features a modified chat_template.jinja that pre-fills <|channel>thought\n on assistant turns when add_generation_prompt=True and enable_thinking=True, addressing a "merges skip thinking" regression and allowing for thought traces.
  • Context Length: Supports a context window of 32768 tokens.

Good For

  • Creative Text Generation: Ideal for applications requiring nuanced and varied creative writing styles.
  • Roleplay and Storytelling: Benefits from the blended styles and improved prompt attention.
  • Research into Merging Techniques: Serves as a V2 data point for instruct-preserving merges, particularly for understanding the impact of Fisher protection and layer-importance damping.

Finetunes
2 models