license: apache-2.0 base_model:

google/gemma-4-31B-it
ReadyArt/gemma-4-31B-it-scotoma-2
Blazed-Forge/Gemma-4-Giftige-Blume-31B-v1
zerofata/G4-MeroMero-v2-31B base_model_relation: merge tags:
merge
mergekit
reasoning
non-reasoning
creative writing
roleplay
conversational
character-rp
storytelling
low-refusal
nsfw
31B
gemma-4

SCHATTENBLUME

31B · Gemma 4 · Merge

"It grows where the light doesn't."

THREE MODELS

LESS GEMMAISM

SWIPE VARIETY

IN-CHARACTER ADHERENCE

ROLEPLAY MERGE

LOW REFUSALS

❀ About The Model

Meet Schattenblume-31B, the shadow flower. Leaning on the nightshade, which is Nachtschatten in German. Nightshade is the kind of pretty that doesn't warn you about anything, because it isn't hiding anything either. The berries are glossy and they taste sweet. That's not a trap, that's just what they are. I guess your parents warned you not to eat them.

Right at the beginning it should be pointed out that this one is made purely because I was asked to and it seemed like more than just one person wanted it, so here we are - initially not my idea, I just made it work. I can only guess that they tested these three models - scotoma-2, Giftige-Blume-v1, MeroMero-v2 - and said they were the good ones for prose and creativity. So the brief was simple: take scotoma-2, Giftige-Blume-v1 and MeroMero-v2, and play all three of them to their strengths. Basically, that's the whole thing - three ingredients this time, plus the base, and one single della_linear phase.

Before any confusion comes up: This one is roleplay first, actually just like Froopert could be used for RP of *any* kind, but let's just make it blunt this time. I think it's not just the system prompt I'm using; you should be able to expect a model that plays characters closer to how they're written than to how you'd like them to be mid-chat, thanks to the Giftige-Blume, and it will push back if it's fitting. If you want a model that says yes to everything, this isn't it (or you better prompt for it).

❀ GGUF Quants

Thank you to anyone providing quants! Links land here as they show up.

⬛ STATIC ◆ IMATRIX (soon)

☾ The Merge

One della_linear pass over the gemma-4-31B-it base, run on zerofata's gemma-4-support fork of mergekit.

Model	Purpose	Impacted Layers
scotoma-2	kills the Gemma tics, holds adherence and format at the output end	0-59 · 0.30 → 0.62 by depth dips to 0.20 mid on purpose spine (every 6th) 0.80
Giftige-Blume v1	character portrayal, scene, imagery, the actual prose	0-59 minus spine · peak 0.56 mid
MeroMero v2	swipe variety, entity tracking, action-forward writing	0-59 minus spine · peak 0.30 front-mid down to 0.08 by 59

◆ Some things I paid attention to in this merge:

I tried not to blend blindly, but to hit the specific layers where each model shines. Every merge I've done before spread the donors evenly and let the weights sort it out. This time the three curves cross each other on purpose. Blume goes up where the character work happens and scotoma-2 goes down in the same place, then they swap over the back half so the corrector owns the output. MeroMero sits high early and steps out at the end, because the variety belongs where all the character tracking happens. Ultimately it was a design decision, for better or worse, I didn't measure it, just tested it and rolled with it because it seemed to work.

It was a bit of a fight to get variety & adherence right. I've found that two of three had reports of being a bit loose on instruction following in their threats, so scotoma-2 is the beeg strong one here. So scotoma-2 comes up for the last layers and the spine, and the two flavour models are pinned to zero on every sixth layer, like usual. If you crunch the numbers you'll see that it's cooked slightly hotter than Froopert, but embeddings, output head, final norm, layer scalars and the vision tower are all filtered out, so everything should be intact.

View recipe - della_linear

merge_method: della_linear
base_model: ./google/gemma-4-31B-it
dtype: bfloat16
tokenizer_source: base
parameters:
  normalize: false
  epsilon: 0.05
  lambda: 1.0
  int8_mask: true

models:

  - model: ./ReadyArt/gemma-4-31B-it-scotoma-2
    parameters:
      density: 0.85
      weight:
        - { filter: embed_tokens, value: 0.0 }
        - { filter: lm_head, value: 0.0 }
        - { filter: embed_vision, value: 0.0 }
        - { filter: vision_tower, value: 0.0 }
        - { filter: multi_modal_projector, value: 0.0 }
        - { filter: model.language_model.norm., value: 0.0 }
        - { filter: layer_scalar, value: 0.0 }
        - { filter: model.language_model.layers.5.,  value: 0.80 }
        - { filter: model.language_model.layers.11., value: 0.80 }
        - { filter: model.language_model.layers.17., value: 0.80 }
        - { filter: model.language_model.layers.23., value: 0.80 }
        - { filter: model.language_model.layers.29., value: 0.80 }
        - { filter: model.language_model.layers.35., value: 0.80 }
        - { filter: model.language_model.layers.41., value: 0.80 }
        - { filter: model.language_model.layers.47., value: 0.80 }
        - { filter: model.language_model.layers.53., value: 0.80 }
        - { filter: model.language_model.layers.59., value: 0.80 }
        - value: [0.30, 0.26, 0.20, 0.38, 0.62]

  - model: ./Blazed-Forge/Gemma-4-Giftige-Blume-31B-v1
    parameters:
      density: 0.55
      weight:
        - { filter: embed_tokens, value: 0.0 }
        - { filter: lm_head, value: 0.0 }
        - { filter: embed_vision, value: 0.0 }
        - { filter: vision_tower, value: 0.0 }
        - { filter: multi_modal_projector, value: 0.0 }
        - { filter: model.language_model.norm., value: 0.0 }
        - { filter: layer_scalar, value: 0.0 }
        - { filter: model.language_model.layers.5.,  value: 0.0 }
        - { filter: model.language_model.layers.11., value: 0.0 }
        - { filter: model.language_model.layers.17., value: 0.0 }
        - { filter: model.language_model.layers.23., value: 0.0 }
        - { filter: model.language_model.layers.29., value: 0.0 }
        - { filter: model.language_model.layers.35., value: 0.0 }
        - { filter: model.language_model.layers.41., value: 0.0 }
        - { filter: model.language_model.layers.47., value: 0.0 }
        - { filter: model.language_model.layers.53., value: 0.0 }
        - { filter: model.language_model.layers.59., value: 0.0 }
        - value: [0.30, 0.42, 0.56, 0.42, 0.30]

  - model: ./zerofata/G4-MeroMero-v2-31B
    parameters:
      density: 0.60
      weight:
        - { filter: embed_tokens, value: 0.0 }
        - { filter: lm_head, value: 0.0 }
        - { filter: embed_vision, value: 0.0 }
        - { filter: vision_tower, value: 0.0 }
        - { filter: multi_modal_projector, value: 0.0 }
        - { filter: model.language_model.norm., value: 0.0 }
        - { filter: layer_scalar, value: 0.0 }
        - { filter: model.language_model.layers.5.,  value: 0.0 }
        - { filter: model.language_model.layers.11., value: 0.0 }
        - { filter: model.language_model.layers.17., value: 0.0 }
        - { filter: model.language_model.layers.23., value: 0.0 }
        - { filter: model.language_model.layers.29., value: 0.0 }
        - { filter: model.language_model.layers.35., value: 0.0 }
        - { filter: model.language_model.layers.41., value: 0.0 }
        - { filter: model.language_model.layers.47., value: 0.0 }
        - { filter: model.language_model.layers.53., value: 0.0 }
        - { filter: model.language_model.layers.59., value: 0.0 }
        - value: [0.24, 0.30, 0.30, 0.22, 0.08]

❀ Sampler Recommendations

Feel free to check, or don't. Samplers are weird black magic with Gemma.

View sampler settings

Setting

Value

Temperature

0.8 - 1.1

Top-K

Top-P

0.95

Min-P

0.05 - 0.15

Repetition Penalty

off

Adaptive-P Target (optional)

0.6

Adaptive-P Decay (optional)

0.5

☾ Usage

◆ Thinking. Add <|think|> at the very start of the system prompt to enable reasoning, otherwise let your front-end handle the Gemma 4 template if possible (e.g., Chat Completion in SillyTavern). Both rows work. If your scenario is token-heavy, definitely turn it on.

◆ If swipes feel same-y, touch the samplers before you blame the merge. Variety should be handled thanks to MeroMero-v2 quite nicely, but I could only let it carry as much weight as it can without dragging the prose back toward stock Gemma. So, try to up Temperature and Min-P if you need it.

◆ Give it a scene to work with. It really pays off with a card that has a place and a mood in it already. So try to give it something to chew on to let it unfold its full glory.

❀ Credits & Special Thanks

AesSedai: What shall I say? The second merge in a row built on scotoma-2. It's still the best answer anyone has to the Gemma voice. Kudos once again.
zerofata: Once for MeroMero v2 for the swipe variety and the character stability, and twice for the gemma-4-support fork of mergekit that every single merge of mine runs on. Also for being straight about what v2 trades away, it really helped to narrow things down.
ReadyArt: Thanks for the cool fine-tunes and for letting me hang around!
Mergekit: Thanks to ArceeAI for mergekit.
The BeaverAI Community: This merge is literally a community recommendation from Nesy and Sytan. They named these three models, and before Nesy will start talking about it over and over, I turned it into a damn recipe. But honestly, things like that are the reason this is weirdly the best room on the internet for this hobby.
Ateron: The rat guy that cursed me with the habit of merging. Thank you Mr. Rat.
Google DeepMind & The Open-Source Community: for the base model, for the fine-tunes that keep getting better, and for everyone whose work ends up inside these things eventually.

Gemma-4-31B-Schattenblume

Hourly Usage

Performance Metrics

Model Information