This merge was made to keep grounded and realistic role-play. Gemma 4 often loses character adherence, slipping its tone into unnecessary lewdness. I tried to push this away while keeping it uncensored. The second focus was on writing stories and adventurous role-play.

Primary Focus: Character adherence. Role-play. Long context stories.

🗣️ Prompt Format

Please refer to the original google/gemma-4-31b-it for the correct chat template.

text

<|turn>user
Hello<turn|>
<|turn>model
Hi there<turn|>
<|turn>user
How are you?<turn|>
<|turn>model

🐀 Merge Details

I used mergekit to combine the specific advantages of models in three-step way:

Phase 1: The Adventure (dare_ties)

First, I took Our Gemsicle Model using it as a base. Then LatitudeGames/Equinox-31B and BirdToast/Gemma-4-31B-glimmer-rp-v0.1- I found them to be good in pro-active writing and role-play, so I carefully blend them together.

dare_ties Recipe 1

merge_method: dare_ties
base_model: F:\AI\Merge\Gemsicle
models:
  - model: F:\AI\Merge\Glimmer
    parameters:
      density: [0.30, 0.40, 0.60, 0.35, 0.40]
      weight: [0.30, 0.35, 0.50, 0.40, 0.30]
  - model: F:\AI\Merge\Equinox
    parameters:
      density: [0.70, 0.60, 0.50, 0.65, 0.60]
      weight: [0.45, 0.45, 0.40, 0.40, 0.50]
parameters:
  lambda: 1.0
  int8_mask: true
dtype: bfloat16

Phase 2: The Tone (dare_ties)

First, I took Our Gemsicle Model again using it as a base. Then Jackrong/Gemopus-4-31B-it and MRockatansky/Gemma-4-31B-storymaxxed2- This one supposed to give coherence in long rp, add more story-genres and keep it uncensored.

dare_ties Recipe 2

merge_method: dare_ties
base_model: F:\AI\Merge\Gemsicle
models:
  - model: F:\AI\Merge\Gemopus
    parameters:
      density: [0.60, 0.45, 0.30, 0.20, 0.20]
      weight: [0.50, 0.30, 0.10, 0.10, 0.10]
  - model: F:\AI\Merge\Gemma-4-31B-storymaxxed2
    parameters:
      density: [0.20, 0.20, 0.40, 0.40, 0.40]
      weight: [0.25, 0.25, 0.50, 0.40, 0.35]
parameters:
  lambda: 1.0
  int8_mask: true
dtype: bfloat16

Phase 3: Final Touch (della_linear)

There I keep rolling with Gemsicle as glued base and put two existing merges, picking the most important layers once more.

della_linear Recipe

merge_method: della_linear
base_model: F:\AI\Merge\Gemsicle
models:
  - model: F:\AI\Merge\SphinsikusPhase1
    parameters:
      density: [0.20, 0.20, 0.50, 0.50, 0.25]
      weight: [0.20, 0.35, 0.50, 0.65, 0.30]
  - model: F:\AI\Merge\SphinsikusPhase2
    parameters:
      density: [0.45, 0.45, 0.30, 0.40, 0.10]
      weight: [0.50, 0.30, 0.30, 0.35, 0.05]
parameters:
  epsilon: 0.08
  lambda: 1.0
  int8_mask: true
  normalize: true
dtype: bfloat16

🤝 Special Thanks

Google DeepMind: For providing the base model.
The Open-Source Community: Creators and all their fine-tunes models that was used in merge.
Mergekit Fork: Zerofata - Thank you for being so nice and making it possible.
To Nimbz: This fella for sharing brain cell.

dr-housemd/Gemma-4-Sphinsikus-Chronist-31B-4.50bpw-exl3

⏳ 📜 Sphinsikus Chronist 📜 ⏳

🐀 About the Model