Instead of traditional linear weight blending—which can dilute specific behavioral fine-tuning and result in generic text—the Model Stock technique maps the fine-tunes as vectors in a geometric space relative to the base model. This mathematical optimization preserves the distinct literary strengths of each component model rather than averaging them out.

🤝 Merged Models

Anthology combines three distinct creative voices to form a singular, cohesive writing instrument:

Component Model	Primary Contribution
MRockatansky/Gemma-4-31B-storymaxxed2	My own tune, which writes great stories from minimal prompting with a writing style close to stock gemma-4, for better or worse.
nbeerbower/Gemma4-Gutenberg-31B-Heretic	Imbues the output with a classic, novel-like prose structure, and a little dash of antiquated vocabulary from the old public domain fiction it was trained on.
llmfan46/gemma-4-Ortenzya-31B-it-uncensored-heretic	Chosen for its vocabulary choices and writing style.

🛠️ Mergekit Configuration

The merge was conducted utilizing zerofata's gemma4-support branch of Mergekit, using the Model Stock merge method.

yaml

merge_method: model_stock
base_model: google/gemma-4-31b-it
dtype: bfloat16
normalize: true
models:
  - model: MRockatansky/Gemma-4-31B-storymaxxed2
  - model: nbeerbower/Gemma4-Gutenberg-31B-Heretic
  - model: llmfan46/gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic

💜 Special Thanks

The Beaver AI Discord community: For my introduction to and education in the wonderful world of local open-source AI
The Tuners: Thanks to nbeerbower and llmfan46 for creating the great models used in this merge
zerofata: For creating the gemma4-support fork of Mergekit that made this merge possible

MRockatansky/gemma-4-31B-anthology

📚 Anthology 31B

📌 Overview

🤝 Merged Models

🛠️ Mergekit Configuration

💜 Special Thanks

MRockatansky/
gemma-4-31B-anthology