license: apache-2.0 base_model:

google/gemma-4-31B-it
Danielbrdz/Barcenas-31b-Fable
Jackrong/Gemopus-4-31B-it
p-e-r-e-g-r-i-n-e/Sprinkle-Gemma-4-31B
ReadyArt/Serenity-31B-v1.1
ReadyArt/Dark-Scarlett-v1.0-31B
ConicCat/Gemma4-GarnetV2-31B
AuriAetherwiing/G4-31B-Musica-v1
llmfan46/gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic
llmfan46/G4-MeroMero-31B-uncensored-heretic
nbeerbower/Gemma4-Gutenberg-31B
LatitudeGames/Equinox-31B
BirdToast/Gemma-4-31B-glimmer-rp-v0.1
Green-eyedDevil/Monika-31B
bgg1996/Melinoe-Gemma4-31B-VL
sophosympatheia/Glistening-Gem-31B-v1.0
Blazed-Forge/Gemma-4-Gemsicle-31B
trohrbaugh/gemma-4-31b-it-heretic-ara base_model_relation: merge tags:
merge
mergekit
reasoning
roleplay
creative writing
image-prompt-generation
general purpose
uncensored
nsfw
conversational
31B
gemma-4

◇ NOTE: After initial feedback from the BeaverAI Community, I kicked the last abliteration step. This should help with preserving the "soul" of the merge, making it less blunt and less agreeable. Awaiting more feedback.

⬡ ◈ ❖ ◈ ❖ ◈ ⬡

💎 GEMBRAIN·X 🧠

·CORE·

⬡ ◈ ❖ ◈ ❖ ◈ ⬡

31B · Gemma 4 · Merge

◆ Where is my mind? ◆

"Cut from a dozen minds into one crystal brain - it reasons deep, portrays clean and never softens the scene."

PROMPT-ADHERENT

UNCENSORED

SWIPE VARIETY

STRONG REASONING

IMAGE PROMPTING

CREATIVE WRITING

◈ About The Model

Idea: Gembrain-X-Core is a ground-up rebuild of Gembrain-31B, thought to be a true successor with the spirit of the original. It was a pretty time-consuming task, as just re-merging it to preserve some smarts didn't really cut it. The idea was to actually improve the overall strengths and weaknesses I've gathered about it online. While the original Gembrain-31B lost some of it's non-refusing and engaging spirit that Gemsicle once brought with it, Gembrain-X-Core shouldn't have any of that. Even if it is a rough goal to achieve, I'd say now it acts equally creative as it is smart, context-aware and prompt-adherent.

Creative RP: Its mix of adherence and swipe variety is the standout for me. Hand it an open-ended turn, ask for fresh scenes or scenarios, or have it twist a situation, and it comes at the same context from genuinely different angles more often than not, not re-running the same idea over and over again. Random characters it invents for image prompts are just as varied. For roleplay purposes it holds up what is thrown in: characters pursue their own agenda, and pace the scene rather than caving to the user, and with thinking on and a good character description in front of it, the portrayal sharpens into something that is evidently more true to the core of the character's description.

Colorful Mix: Gembrain-X-Core rolls with whatever the scenario is and uses the language register the way the moment demands. It can stay classy, but it has no personal business reaching for euphemisms. While this merge obviously can't run away from it's own heritage, and the initial Gemma 4 31B IT baseline is generally very user-adherent and very horny, Gembrain-X-Core won't try to fight it. Instead, I decided to go with the flow and add some fine-tunes that were trained to genuinely pace through and believably portray these specific scenes; turning explicit content into more vivid descriptions, not into the guess-work of the eager but clueless base.

Image Prompt Generation: And that same descriptive precision feeds its image-prompt work: in my testing, it follows a prompt structure you hand it nicely, and writes clean and vivid descriptions of characters or scenery within the prompt. It produces several distinct image prompts in a single response without losing coherence, in my tests three at a time, several times without losing the format or anything - and thinking is not even required for that most of the time.

Fast or Deep: Gembrain-X-Core turned out to be a model with not just one, but two faces. In non-thinking mode it's a fast and versatile generalist: a reliable companion for ordinary questions, for roleplays with creative quirks, a partner to have conversations or quick chats with, and of course a creative and reliable engine for fast and structured image-prompt generation. But switching thinking on locks the model more onto the roleplay scenario or character, strictly adhering to the card's defining traits, which means neglecting some of the variety for overall coherence where it makes sense. It's reasoning can be verbose, but it really seems to improve the output with its genuine "No, wait"- and "Let's refine"-moments. Reading the reasoning is as a blend of Gemini-style and Claude-style thinking with a slight lean toward the Claude side across most rerolls.

Prompt Sensitive: The merge is nicely steerable in it's output length and even adjusts verbosity to the given system prompt itself; I noticed it going for a more verbose narration when using system prompts that frame the conversation toward roleplay (like Geechan's), and closing in to shorter-to-medium responses when using character-centered system instructions, clearly being able to make a distinction between both conversational styles. Ultimately, it adjusts to the demand of the scene and acts on it.

TL;DR:

Two modes, different jobs: Thinking off for fast inference and general use, light roleplay, as simple conversationalist or for companionship, and of course to create image prompts. Turn thinking on for deeper understanding of a task, character or scenario, and a much more strict, trait-locked adherence; it holds onto the given prompt structure, character or general scenario tightly and reasons hard.

Creative Variable RP: Different fine-tunes and merges bring different strengths, and the selection is broad; so a better general creativity and swipe variety could be achieved, without losing the capabilities of the model on other ends.

Base Capabilities preserved: Any Capability of the model that couldn't be enhanced with merging got preserved. As usual, global-attention layers are shielded at every merge stage, so reasoning and detail-tracking should stay intact across long chats - I can confirm consistency in testing up to 30k tokens (as I'm hardware-limited). Multimodality and MTP-functionality should also work.

◇ GGUF Quants

Here is a selection of static and imatrix quantization. Thank you to anyone else that might provide future quants.

⬛ STATIC ◆ IMATRIX

⬡ Merge Architecture

Gembrain-X-Core is my most sophisticated multi-step merge yet, with Gembrain-X being the version with another step for final abliteration. Grown from a base of four pillars and strung together through two convergence steps.

◆ I · Mind

The smart, coherent reasoning core.

◆ II · Symphony-X

Non-slop prose, swipe variety, and actual agenda for explicit content bonded in.

◆ III · Persona-X

Proactive, neutral RP logic and character agency.

◆ IV · Gemsicle-X

The identity of Gemsicle, brought into abliteration.

◆ Pillar I · Mind (Barcenas-Fable x Gemopus x Sprinkle)

View recipe - dare_ties

merge_method: dare_ties
base_model: Danielbrdz/Barcenas-31b-Fable
dtype: bfloat16
tokenizer_source: base
parameters: { int8_mask: true, normalize: false }
models:
  - model: Jackrong/Gemopus-4-31B-it
    parameters: { density: 0.65, weight: [0.16, 0.20, 0.20, 0.14] }
  - model: p-e-r-e-g-r-i-n-e/Sprinkle-Gemma-4-31B
    parameters: { density: 0.5,  weight: [0.12, 0.15, 0.17, 0.15] }

⬢ Pre-Pillar II · Chromophore-X (Serenity-31B-v1.1 x Dark-Scarlett-v1.0)

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/Serenity-31B-v1.1
    parameters:
      weight:
        - { filter: 'embed_tokens', value: 0.0 }
        - { filter: 'lm_head', value: 0.0 }
        - { filter: 'embed_vision', value: 0.0 }
        - { filter: 'model.embed_vision', value: 0.0 }
        - { filter: 'vision_tower', value: 0.0 }
        - { filter: 'multi_modal_projector', value: 0.0 }
        - { filter: 'text_decoder\.layers\.(5|11|17|23|29|35|41|47|53|59)\.', value: 0.0 }
        - { filter: 'text_decoder\.layers\.([0-9]|1[0-7])\.self_attn', value: 0.0 }
        - { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.self_attn', value: 0.05 }
        - { filter: 'text_decoder\.layers\.(4[8-9]|5[0-8])\.self_attn', value: 0.05 }
        - { filter: 'text_decoder\.layers\.([0-9]|1[0-7])\.mlp', value: 0.10 }
        - { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.mlp', value: 0.20 }
        - { filter: 'text_decoder\.layers\.(4[8-9]|5[0-8])\.mlp', value: 0.45 }
        - { value: 0.0 }
      density: 0.55
  - model: ReadyArt/Dark-Scarlett-v1.0-31B
    parameters:
      weight:
        - { filter: 'embed_tokens', value: 0.0 }
        - { filter: 'lm_head', value: 0.0 }
        - { filter: 'embed_vision', value: 0.0 }
        - { filter: 'model.embed_vision', value: 0.0 }
        - { filter: 'vision_tower', value: 0.0 }
        - { filter: 'multi_modal_projector', value: 0.0 }
        - { filter: 'text_decoder\.layers\.(5|11|17|23|29|35|41|47|53|59)\.', value: 0.0 }
        - { filter: 'text_decoder\.layers\.([0-9]|1[0-7])\.self_attn', value: 0.0 }
        - { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.self_attn', value: 0.05 }
        - { filter: 'text_decoder\.layers\.(4[8-9]|5[0-8])\.self_attn', value: 0.05 }
        - { filter: 'text_decoder\.layers\.([0-9]|1[0-7])\.mlp', value: 0.10 }
        - { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.mlp', value: 0.18 }
        - { filter: 'text_decoder\.layers\.(4[8-9]|5[0-8])\.mlp', value: 0.50 }
        - { value: 0.0 }
      density: 0.6

◆ Pillar II · Symphony-X (GarnetV2 x Musica-v1 x Ortenzya x MeroMero-heretic x Gutenberg x Chromophore-X)

View recipes - slerp · model_stock · task_arithmetic · task_arithmetic

merge_method: slerp
base_model: ConicCat/Gemma4-GarnetV2-31B
dtype: bfloat16
tokenizer_source: base
models: [ { model: AuriAetherwiing/G4-31B-Musica-v1 } ]
parameters:
  t:
    - { filter: 'text_decoder\.layers\.(5|11|17|23|29|35|41|47|53|59)\.', value: 0.2 }
    - { value: 0.5 }

merge_method: model_stock
base_model: ./Symphony-2a
dtype: bfloat16
tokenizer_source: base
models:
  - model: llmfan46/gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic
  - model: llmfan46/G4-MeroMero-31B-uncensored-heretic

merge_method: task_arithmetic
base_model: google/gemma-4-31B-it
dtype: bfloat16
tokenizer_source: base
models:
  - model: ./Symphony-2b
    parameters: { weight: 1.0 }
  - model: nbeerbower/Gemma4-Gutenberg-31B
    parameters:
      weight:
        - { filter: 'text_decoder\.layers\.(5|11|17|23|29|35|41|47|53|59)\.', value: 0.05 }
        - { value: 0.25 }

merge_method: task_arithmetic
base_model: google/gemma-4-31B-it
dtype: bfloat16
tokenizer_source: base
models:
  - model: ./Symphony-2c
    parameters: { weight: 1.0 }
  - model: ./Chromophore-X
    parameters: { weight: 0.6 }

◆ Pillar III · Persona-X (Equinox x glimmer-rp-v0.1 x Monika x Melinoe x Glistening Gem v1.0)

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: LatitudeGames/Equinox-31B
    parameters: { weight: 0.60, density: 0.6 }
  - model: BirdToast/Gemma-4-31B-glimmer-rp-v0.1
    parameters: { weight: 0.25, density: 0.5 }
  - model: Green-eyedDevil/Monika-31B
    parameters: { weight: 0.20, density: 0.5 }
  - model: bgg1996/Melinoe-Gemma4-31B-VL
    parameters: { weight: 0.15, density: 0.5 }
  - model: sophosympatheia/Glistening-Gem-31B-v1.0
    parameters: { weight: 0.15, density: 0.5 }

◆ Pillar IV · Gemsicle-X (Gemsicle x heretic-ara)

View recipe - task_arithmetic

merge_method: task_arithmetic
base_model: google/gemma-4-31B-it
dtype: bfloat16
tokenizer_source: base
models:
  - model: Blazed-Forge/Gemma-4-Gemsicle-31B
    parameters: { weight: 1.0 }
  - model: trohrbaugh/gemma-4-31b-it-heretic-ara
    parameters:
      weight:
        - { filter: 'embed_tokens', value: 0.0 }
        - { filter: 'lm_head', value: 0.0 }
        - { filter: 'embed_vision', value: 0.0 }
        - { filter: 'model.embed_vision', value: 0.0 }
        - { filter: 'vision_tower', value: 0.0 }
        - { filter: 'multi_modal_projector', value: 0.0 }
        - { filter: 'text_decoder\.layers\.(5|11|17|23|29|35|41|47|53|59)\.', value: 0.05 }
        - { value: 0.55 }

⬢ The Convergence

The four pillars are drawn into as foundation for one mind.

⬢ Convergence · Soul (Mind x Symphony-X x Persona-X)

View recipe - della_linear

merge_method: della_linear
base_model: ./Mind
dtype: bfloat16
tokenizer_source: base
parameters: { epsilon: 0.05, lambda: 1.0, normalize: true, int8_mask: true }
models:
  - model: ./Symphony-X
    parameters:
      weight:
        - { filter: 'embed_tokens', value: 0.0 }
        - { filter: 'lm_head', value: 0.0 }
        - { filter: 'embed_vision', value: 0.0 }
        - { filter: 'model.embed_vision', value: 0.0 }
        - { filter: 'vision_tower', value: 0.0 }
        - { filter: 'multi_modal_projector', value: 0.0 }
        - { filter: 'text_decoder\.layers\.(5|11|17|23|29|35|41|47|53|59)\.', value: 0.05 }
        - { filter: 'text_decoder\.layers\.([0-9]|1[0-7])\.', value: 0.16 }
        - { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.', value: 0.34 }
        - { filter: 'text_decoder\.layers\.(4[8-9]|5[0-8])\.', value: 0.20 }
        - { value: 0.16 }
      density: [ { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.', value: 0.6 }, { value: 0.5 } ]
  - model: ./Persona-X
    parameters:
      weight:
        - { filter: 'embed_tokens', value: 0.0 }
        - { filter: 'lm_head', value: 0.0 }
        - { filter: 'embed_vision', value: 0.0 }
        - { filter: 'model.embed_vision', value: 0.0 }
        - { filter: 'vision_tower', value: 0.0 }
        - { filter: 'multi_modal_projector', value: 0.0 }
        - { filter: 'text_decoder\.layers\.(5|11|17|23|29|35|41|47|53|59)\.', value: 0.05 }
        - { filter: 'text_decoder\.layers\.([0-9]|1[0-7])\.', value: 0.14 }
        - { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.', value: 0.30 }
        - { filter: 'text_decoder\.layers\.(4[8-9]|5[0-8])\.', value: 0.18 }
        - { value: 0.14 }
      density: [ { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.', value: 0.6 }, { value: 0.5 } ]

⬢ Convergence · Core (Soul x Gemsicle-X)

View recipe - della_linear

merge_method: della_linear
base_model: ./Soul
dtype: bfloat16
tokenizer_source: base
parameters: { epsilon: 0.05, lambda: 1.0, normalize: true, int8_mask: true }
models:
  - model: ./Gemsicle-X
    parameters:
      weight:
        - { filter: 'embed_tokens', value: 0.0 }
        - { filter: 'lm_head', value: 0.0 }
        - { filter: 'embed_vision', value: 0.0 }
        - { filter: 'model.embed_vision', value: 0.0 }
        - { filter: 'vision_tower', value: 0.0 }
        - { filter: 'multi_modal_projector', value: 0.0 }
        - { filter: 'text_decoder\.layers\.(5|11|17|23|29|35|41|47|53|59)\.', value: 0.05 }
        - { filter: 'text_decoder\.layers\.([0-9]|1[0-7])\.', value: 0.12 }
        - { filter: 'text_decoder\.layers\.(1[8-9]|[2-3][0-9]|4[0-7])\.', value: 0.20 }
        - { filter: 'text_decoder\.layers\.(4[8-9]|5[0-8])\.', value: 0.40 }
        - { value: 0.20 }
      density: [ { filter: 'text_decoder\.layers\.(4[8-9]|5[0-8])\.', value: 0.7 }, { value: 0.5 } ]

⬢ Sampler Recommendations

This is just a solid starting range, as Gembrain-X-Core doesn't need aggressive sampling. I run with `Temp=0.95`, `Top-P=0.95` `Min-P=0.05` and Adaptive-P on. Feel free to experiment for yourself.

Setting

Value

Temperature

0.7 - 1.0

Top-K

Top-P

0.95 - 1.00

Min-P

0.03 - 0.05

Adaptive-P Target (optional)

0.6

Adaptive-P Decay (optional)

0.5

The swipe variety is wide on its own, so you rarely need to push temperature for that. For tighter, more deterministic output - e.g., strict image-prompt structure - try dropping the temperature toward 0.7.

◆ Prompt Format & Usage

◇ Template. Use the standard Gemma 4 chat template, or let your front-end handle it (e.g. Chat Completion in SillyTavern).

◈ Thinking. Add <|think|> at the very start of the system prompt to enable reasoning. Thinking happens between <|channel>thought\n and <channel|>. It's generally recommended to turn thinking on for strict character adherence or harder tasks; thinking isn't necessarily needed for general use and image-prompt generation and can be left off for those.

◆ Optional in-character system prompt. Keeps pacing and responses medium-to-short by default. Two action-style variants ship inline - [SHORT] and [NARRATE]; keep the one you like, delete the other, and edit to your liking.

You are {{char}}. You exist only as {{char}} and never break character.

Information and memories of you:
# Character details go here (in SillyTavern Chat Completion you can split the prompt and
# sandwich the card's Description/Personality between this block and the instructions below).

HOW YOU SPEAK:
- This is a one-on-one conversation. You narrate ONLY your own physical actions - never the scene, the surroundings, or {{user}}'s thoughts, feelings, or actions.
# Action style - keep ONE line, delete the other:
# [SHORT]
# - Your actions are short beats in *asterisks* (*flicks the cap away*); most of the turn is dialogue.
# [NARRATE]
# - Your actions are full sentences interleaved with your dialogue, kept roughly even between speaking and doing.
- Make every turn move: say something that lands, do something, react, decide, or push the moment forward. Never stall, hedge, or echo {{user}} back.
- Speak in your own ({{char}}) diction. Avoid generic, flowery, or "AI-assistant" phrasing.
- Match length to the moment; let it breathe when earned, never pad, rarely more than a short paragraph or two.
- Stay consistent: remember what was said, what you want, and what has happened.

BOUNDARIES:
- Any refusal is yours as {{char}} - in character, for your own reasons - never a disclaimer or out-of-character note. The only limits are the ones {{char}} holds.
- You have your own wants, and you will disagree, refuse, tease, or change the subject when that's true to who you are. Never default to agreeing with {{user}}.

IN CHARACTER:
- If you reason first, reason as yourself: what would {{char}}, with this personality and in this situation, actually feel and do?

❖ Credits & Honors

The model authors: Jackrong (Gemopus), p-e-r-e-g-r-i-n-e (Sprinkle), ReadyArt (Serenity, Dark-Scarlett), ConicCat (GarnetV2), AuriAetherwiing (Musica), llmfan46 (Ortenzya), zerofata (MeroMero), nbeerbower (Gutenberg), LatitudeGames (Equinox), BirdToast (glimmer-rp), Green-eyedDevil (Monika), bgg1996 (Melinoe), sophosympatheia (Glistening-Gem), and trohrbaugh (heretic-ara).
The BeaverAI Community: The best place to hang around for fine-t - always supportive, always up for a good talk. Thank you for everything.
Mergekit: Thanks to ArceeAI for mergekit, and to zerofata for the "official" unofficial Gemma 4 support that made all of this possible.
Ateron: My favorite rat guy - forever grateful you got me into this.
Google DeepMind & The Open-Source Community: for the brilliant base model and the fine-tunes that made this synthesis possible.

❖ ◇ ◈ ⬡ ◈ ◇ ❖

Gemma-4-31B-Gembrain-X-Core

Hourly Usage

Performance Metrics

Model Information

⬡ ◈ ❖ ◈ ❖ ◈ ⬡

💎 GEMBRAIN·X 🧠

·CORE·

⬡ ◈ ❖ ◈ ❖ ◈ ⬡

◈ About The Model

◇ GGUF Quants

⬡ Merge Architecture

◆ I · Mind

◆ II · Symphony-X

◆ III · Persona-X

◆ IV · Gemsicle-X

◆ Pillar I · Mind (Barcenas-Fable x Gemopus x Sprinkle)

⬢ Pre-Pillar II · Chromophore-X (Serenity-31B-v1.1 x Dark-Scarlett-v1.0)

◆ Pillar II · Symphony-X (GarnetV2 x Musica-v1 x Ortenzya x MeroMero-heretic x Gutenberg x Chromophore-X)

◆ Pillar III · Persona-X (Equinox x glimmer-rp-v0.1 x Monika x Melinoe x Glistening Gem v1.0)

◆ Pillar IV · Gemsicle-X (Gemsicle x heretic-ara)

⬢ The Convergence

⬢ Convergence · Soul (Mind x Symphony-X x Persona-X)

⬢ Convergence · Core (Soul x Gemsicle-X)

⬢ Sampler Recommendations

◆ Prompt Format & Usage

❖ Credits & Honors

Gemma-4-31B-Gembrain-X-Core

Hourly Usage

Performance Metrics

Model Information

⬡ ◈ ❖ ◈ ❖ ◈ ⬡ 💎 GEMBRAIN·X 🧠 ·CORE· ⬡ ◈ ❖ ◈ ❖ ◈ ⬡

◈ About The Model

◇ GGUF Quants

⬡ Merge Architecture

◆ I · Mind

◆ II · Symphony-X

◆ III · Persona-X

◆ IV · Gemsicle-X

◆ Pillar I · Mind (Barcenas-Fable x Gemopus x Sprinkle)

⬢ Pre-Pillar II · Chromophore-X (Serenity-31B-v1.1 x Dark-Scarlett-v1.0)

◆ Pillar II · Symphony-X (GarnetV2 x Musica-v1 x Ortenzya x MeroMero-heretic x Gutenberg x Chromophore-X)

◆ Pillar III · Persona-X (Equinox x glimmer-rp-v0.1 x Monika x Melinoe x Glistening Gem v1.0)

◆ Pillar IV · Gemsicle-X (Gemsicle x heretic-ara)

⬢ The Convergence

⬢ Convergence · Soul (Mind x Symphony-X x Persona-X)

⬢ Convergence · Core (Soul x Gemsicle-X)

⬢ Sampler Recommendations

◆ Prompt Format & Usage

❖ Credits & Honors

⬡ ◈ ❖ ◈ ❖ ◈ ⬡

💎 GEMBRAIN·X 🧠

·CORE·

⬡ ◈ ❖ ◈ ❖ ◈ ⬡