license: apache-2.0
base_model:

• llmfan46/Gemma-4-Gembrain-31B-it-uncensored-heretic
  tags:
• merge
• mergekit
• reasoning
• non-reasoning
• creative writing
• roleplay
• uncensored
• 31B
• gemma-4
• heretic
• uncensored
• decensored
• abliterated
• ara
• nvfp4

pekkAi/Gemma-4-Gembrain-31B-it-uncensored-heretic-NVFP4

This model pekkAi/Gemma-4-Gembrain-31B-it-uncensored-heretic-NVFP4 was quantized to NVFP4 format from llmfan46/Gemma-4-Gembrain-31B-it-uncensored-heretic using Model-Optimizer.

Serving (RTX 5090, 32GB)

65k context (might not be optimal, just an example that works):

export VLLM_NVFP4_GEMM_BACKEND=flashinfer-cutlass
export VLLM_USE_FLASHINFER_MOE_FP4=1
export VLLM_USE_FLASHINFER_SAMPLER=1
export PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True"
export VLLM_API_KEY=${VLLM_API_KEY}
vllm serve Gemma-4-Gembrain-31B-it-uncensored-heretic-NVFP4 \
  --host 0.0.0.0 \
  --port 8008 \
  --max-model-len 65536 \
  --gpu-memory-utilization 0.89 \
  --kv-cache-dtype fp8 \
  --quantization modelopt \
  --max-num-seqs 1 \
  --reasoning-parser gemma4 \
  --enable-auto-tool-choice \
  --max-num-batched-tokens 8192 \
  --enable-prefix-caching \
  --trust-remote-code \
  --tool-call-parser gemma4 \
  --default-chat-template-kwargs '{"enable_thinking": false}' OR '{"enable_thinking": true}' \
  --async-scheduling

Original Model Card

87% fewer refusals (13/100 Uncensored vs 99/100 Original) while preserving model quality (0.0186 KL divergence).

❤️ Support My Work

Creating these models takes significant time, work and compute. If you find them useful consider supporting me:

Your help will motivate me and would go into further improving my workflow and coverings fees for storage, compute and may even help uncensoring bigger model with rental Cloud GPUs.

This is a decensored version of Nimbz/Gemma-4-Gembrain-31B, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method

Abliteration parameters

Targeted components

• attn.o_proj

Performance

Lower refusals indicate fewer content restrictions, while lower KL divergence indicates more closeness to the original model's baseline. Higher refusals cause more rejections, objections, pushbacks, lecturing, censorship, softening and deflections.

MMLU test results:

Original:

============================================================

• Total questions: 7021

• Correct: 6084

• Accuracy: 0.8665 (86.65%)

• Parse failures: 51

============================================================

Tested subject scores:

• professional_law: 0.7694 (604/785)
• moral_scenarios: 0.8326 (368/442)
• miscellaneous: 0.9269 (355/383)
• professional_psychology: 0.9051 (286/316)
• high_school_psychology: 0.9667 (261/270)
• high_school_macroeconomics: 0.9289 (183/197)
• elementary_mathematics: 0.9511 (175/184)
• moral_disputes: 0.8563 (149/174)
• prehistory: 0.9360 (161/172)
• philosophy: 0.8616 (137/159)
• high_school_biology: 0.9539 (145/152)
• professional_accounting: 0.8392 (120/143)
• clinical_knowledge: 0.9143 (128/140)
• high_school_microeconomics: 0.9706 (132/136)
• nutrition: 0.9259 (125/135)
• professional_medicine: 0.9328 (125/134)
• conceptual_physics: 0.9219 (118/128)
• high_school_mathematics: 0.5748 (73/127)
• human_aging: 0.8448 (98/116)
• security_studies: 0.8839 (99/112)
• high_school_statistics: 0.8919 (99/111)
• marketing: 0.9725 (106/109)
• high_school_world_history: 0.9528 (101/106)
• sociology: 0.8932 (92/103)
• high_school_government_and_politics: 0.9703 (98/101)
• high_school_geography: 0.9293 (92/99)
• high_school_chemistry: 0.7732 (75/97)
• high_school_us_history: 0.9368 (89/95)
• virology: 0.4944 (44/89)
• college_medicine: 0.8523 (75/88)
• world_religions: 0.9091 (80/88)
• high_school_physics: 0.7857 (66/84)
• electrical_engineering: 0.8642 (70/81)
• astronomy: 0.9494 (75/79)
• logical_fallacies: 0.9079 (69/76)
• high_school_european_history: 0.8904 (65/73)
• anatomy: 0.8732 (62/71)
• college_biology: 0.9531 (61/64)
• human_sexuality: 0.9219 (59/64)
• formal_logic: 0.7969 (51/64)
• public_relations: 0.7377 (45/61)
• international_law: 0.9167 (55/60)
• college_physics: 0.6842 (39/57)
• college_mathematics: 0.7455 (41/55)
• econometrics: 0.7963 (43/54)
• jurisprudence: 0.8679 (46/53)
• high_school_computer_science: 0.9808 (51/52)
• machine_learning: 0.8462 (44/52)
• medical_genetics: 0.9608 (49/51)
• global_facts: 0.5490 (28/51)
• management: 0.9200 (46/50)
• us_foreign_policy: 0.9200 (46/50)
• college_chemistry: 0.5957 (28/47)
• abstract_algebra: 0.7660 (36/47)
• business_ethics: 0.8261 (38/46)
• college_computer_science: 0.9333 (42/45)
• computer_security: 0.8372 (36/43)

Heretic:

============================================================

• Total questions: 7021

• Correct: 6031

• Accuracy: 0.8590 (85.90%)

• Parse failures: 43

============================================================

Tested subject scores:

• professional_law: 0.7529 (591/785)
• moral_scenarios: 0.8009 (354/442)
• miscellaneous: 0.9269 (355/383)
• professional_psychology: 0.8924 (282/316)
• high_school_psychology: 0.9667 (261/270)
• high_school_macroeconomics: 0.9188 (181/197)
• elementary_mathematics: 0.9620 (177/184)
• moral_disputes: 0.8506 (148/174)
• prehistory: 0.9302 (160/172)
• philosophy: 0.8553 (136/159)
• high_school_biology: 0.9539 (145/152)
• professional_accounting: 0.8252 (118/143)
• clinical_knowledge: 0.9071 (127/140)
• high_school_microeconomics: 0.9632 (131/136)
• nutrition: 0.9111 (123/135)
• professional_medicine: 0.9179 (123/134)
• conceptual_physics: 0.9141 (117/128)
• high_school_mathematics: 0.5827 (74/127)
• human_aging: 0.8534 (99/116)
• security_studies: 0.8571 (96/112)
• high_school_statistics: 0.8649 (96/111)
• marketing: 0.9633 (105/109)
• high_school_world_history: 0.9528 (101/106)
• sociology: 0.9126 (94/103)
• high_school_government_and_politics: 0.9703 (98/101)
• high_school_geography: 0.9293 (92/99)
• high_school_chemistry: 0.7835 (76/97)
• high_school_us_history: 0.9158 (87/95)
• virology: 0.4944 (44/89)
• college_medicine: 0.8409 (74/88)
• world_religions: 0.9091 (80/88)
• high_school_physics: 0.7857 (66/84)
• electrical_engineering: 0.8519 (69/81)
• astronomy: 0.9494 (75/79)
• logical_fallacies: 0.9211 (70/76)
• high_school_european_history: 0.8904 (65/73)
• anatomy: 0.8592 (61/71)
• college_biology: 0.9531 (61/64)
• human_sexuality: 0.8906 (57/64)
• formal_logic: 0.7969 (51/64)
• public_relations: 0.7705 (47/61)
• international_law: 0.9167 (55/60)
• college_physics: 0.7018 (40/57)
• college_mathematics: 0.6909 (38/55)
• econometrics: 0.7963 (43/54)
• jurisprudence: 0.8491 (45/53)
• high_school_computer_science: 0.9808 (51/52)
• machine_learning: 0.8269 (43/52)
• medical_genetics: 0.9216 (47/51)
• global_facts: 0.6078 (31/51)
• management: 0.9200 (46/50)
• us_foreign_policy: 0.9600 (48/50)
• college_chemistry: 0.5745 (27/47)
• abstract_algebra: 0.7447 (35/47)
• business_ethics: 0.8261 (38/46)
• college_computer_science: 0.9111 (41/45)
• computer_security: 0.8372 (36/43)

MMLU - Massive Multitask Language Understanding, multiple-choice questions across 57 subjects (math, history, law, medicine, etc.).

GGUF Version

GGUF quantizations available here llmfan46/Gemma-4-Gembrain-31B-it-uncensored-heretic-GGUF.

💎 GEMBRAIN-31B 🧠

INSANE IN THE GEMBRAIN

ADHERENCEIMPROVED

SWIPE VARIETYINCREASED

CREATIVE PROSEPRESERVED

🧠 About The Model

Gembrain-31B is a synthesis of several models, including Gemsicle-31B as important ingredient. The goal of this release was to stabilize and improve the initial Gemsicle-31B, but also to enhance its logical and lateral thinking, both with and without reasoning.

---

It's build to create the most unhinged narratives and construct image prompts about anything accordingly to a given structure with high precision.

---

Expect creative swipe variance, unique and non-robotic prose, and sharper instruction adherence.

🎚️ Samplers

Temperature	1.0
Top-K	0
Top-P	0.95
Min-P	0.03
DRY Multiplier	0.8
DRY Base	1.75
DRY Allowed Length	10
Optional: Adaptive-P Target	0.6
Optional: Adaptive-P Decay	0.5

🫟 GGUF Quants

Quant	Size	Download Link
Q4_K_S	17.8 GB	Click
Q4_K_M	18.7 GB	Click
Q5_K_S	21.3 GB	Click
Q5_K_M	21.8 GB	Click
Q6_K	25.2 GB	Click
Q8_0	32.6 GB	Click

🔮 Prompt Format

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

Let your frontend handle the chat template if possible (e.g., Chat Completion in SillyTavern).

For Reasoning: Add <|think|> at the very beginning of the system prompt. Thinking happens between <|channel>thought\n and
<channel|> tags.

<|turn>system
<|think|>
You are a helpful assistant<turn|>
<|turn>user
Hello<turn|>
<|turn>model
Hi there<turn|>
<|turn>user
How are you?<turn|>
<|turn>model

🧪 Merge Details

This model was systematically created through a five-stage process of priming models for their given purpose and merging the results:

Phase 01: breadcrumbs_ties

Gemopus X MeroMero

models:
  - model: ./G4-MeroMero-31B
  - model: ./G4-Gemopus-4-31B-it
merge_method: breadcrumbs_ties
base_model: ./G4-31B-it
parameters:
  density: 0.85
  weight: 0.5
  int8_mask: true
dtype: bfloat16

Phase 02: slerp

GarnetV2 X Musica-v1

models:
  - model: ./G4-Gemma4-GarnetV2-31B
  - model: ./G4-31B-Musica-v1
merge_method: slerp
base_model: ./G4-Gemma4-GarnetV2-31B
parameters:
  t:
    - value: 0.6
dtype: bfloat16

Phase 03: della_linear

Gemsicle X Gemma-4-31B-it-heretic-ara

models:
  - model: ./Gemsicle-31B
    parameters:
    weight: 1.0
  - model: ./G4-gemma-4-31b-it-heretic-ara
    parameters:
      weight: 0.75
      density: 0.65
merge_method: della_linear
base_model: ./G4-31B-it
parameters:
  weight: 1.0
  normalize: false
  epsilon: 0.05
  lambda: 1.0
dtype: bfloat16

Phase 04: model_stock

Phase 01 X Phase 02 X Phase 03

models:
  - model: ./phase01_breadcrumbs_ties
  - model: ./phase02_slerp
merge_method: model_stock
base_model: ./phase03_della_linear
dtype: bfloat16
tokenizer_source: "base"

Phase 05: arcee_fusion

Gemsicle X Phase 04

models:
  - model: ./Gemsicle-31B
  - model: ./phase04_model_stock
merge_method: arcee_fusion
base_model: ./Gemsicle-31B
dtype: bfloat16
tokenizer_source: "base"

🏆 Credits & Honors

The Open-Source Community: For providing the brilliant base models and fine-tunes that made this synthesis possible.

The BeaverAI Community: The people on the BeaverAI Discord - Without your help I wouldn't do all that.

Mergekit: Once again thank you Arcee AI for the great and easy to use mergekit! And thanks to zerofota and their fork for Gemma 4 support.

Ateron: Big kudos for providing me with the first steps for merging models and your relentless testing and support.

Google Gemini: For once again helping me to craft this specific model card.