Blazed-Forge/Gemma-4-Gemsicle-31B

VISIONConcurrency Cost:2Model Size:31BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 2, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Gemma-4-Gemsicle-31B is a 31 billion parameter experimental model merge by Blazed-Forge, built upon the Google Gemma 4 31B base model with a 32768 token context length. This model is specifically designed for creative writing, aiming to produce less stereotypical character portrayals and more varied prose while maintaining intelligence and prompt adherence. It combines the strengths of highly creative fine-tunes with the base model's intelligence through a two-phase merge process using SLERP and DARE-TIES methods.

Loading preview...

What the fuck is this model about?

Gemma-4-Gemsicle-31B is an experimental 31 billion parameter model merge by Blazed-Forge, developed by Ateron and Nimbz. It is built on the google/gemma-4-31b-it base model, aiming to enhance its creative writing capabilities without sacrificing intelligence or prompt adherence. The model addresses the base Gemma 4's tendency for stereotypical character portrayals and limited narrative variety.

What makes THIS different from all the other models?

This model stands out due to its unique two-phase merging process, combining specific fine-tunes to achieve a balance between creativity and intelligence. It uses Spherical Linear Interpolation (SLERP) to merge AuriAetherwiing/G4-31B-Musica-v1 and ConicCat/Gemma4-GarnetV2-31B for creative prose, followed by a DARE-TIES merge with p-e-r-e-g-r-i-n-e/Sprinkle-Gemma-4-31B to preserve intelligence and prompt adherence. This results in a model optimized for authentic character portrayal and diverse narrative paths.

Should I use this for my use case?

Key Capabilities:

  • Creative Writing: Optimized for generating varied and imaginative prose.
  • Authentic Character Portrayal: Designed to avoid stereotypical character representations.
  • Prompt Adherence: Maintains the ability to follow user instructions accurately.
  • Reduced "Slop" Phrases: Aims for higher quality and less generic output.

Good for:

  • Applications requiring rich, non-formulaic creative text generation.
  • Interactive storytelling, role-playing, and character-driven narratives.
  • Users seeking a model that balances advanced intelligence with enhanced creativity.