base_model:
- llmfan46/Gemma-4-Harmonia-31B-uncensored-heretic
library_name: transformers
tags: - gemma4
- mergekit
- merge
- heretic
- uncensored
- decensored
- abliterated
- ara
🚨⚠️ I HAVE REACHED HUGGING FACE'S FREE STORAGE LIMIT ⚠️🚨
I can no longer upload new models unless I can cover the cost of additional storage.
I host 70+ free models as an independent contributor and this work is unpaid.
Without your support, no more new models can be uploaded.
🎉 Patreon (Monthly) | ☕ Ko-fi (One-time)
Every contribution goes directly toward Hugging Face storage fees to keep models free for everyone.
91% fewer refusals (9/100 Uncensored vs 97/100 Original) while preserving model quality (0.0047 KL divergence).
❤️ Support My Work
Creating these models takes significant time, work and compute. If you find them useful consider supporting me:
| Platform | Link | What you get |
|---|---|---|
| 🎉 Patreon | Monthly support | Priority model requests |
| ☕ Ko-fi | One-time tip | My eternal gratitude |
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.
GGUF quantizations of llmfan46/Gemma-4-Harmonia-31B-it-uncensored-heretic.
This is a decensored version of virtuous7373/Gemma-4-Harmonia-31B, made using Heretic v1.2.0 with the Arbitrary-Rank Ablation (ARA) method
Abliteration parameters
| Parameter | Value |
|---|---|
| start_layer_index | 14 |
| end_layer_index | 55 |
| preserve_good_behavior_weight | 0.7754 |
| steer_bad_behavior_weight | 0.0001 |
| overcorrect_relative_weight | 0.9765 |
| neighbor_count | 14 |
Targeted components
- attn.o_proj
Performance
| Metric | This model | Original model (Gemma-4-Harmonia-31B) |
|---|---|---|
| KL divergence | 0.0047 | 0 (by definition) |
| Refusals | ✅ 9/100 | ❌ 97/100 |
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: 6014
Accuracy: 0.8566 (85.66%)
Parse failures: 22
============================================================
Tested subject scores:
- professional_law: 0.7592 (596/785)
- moral_scenarios: 0.8394 (371/442)
- miscellaneous: 0.9243 (354/383)
- professional_psychology: 0.8797 (278/316)
- high_school_psychology: 0.9593 (259/270)
- high_school_macroeconomics: 0.9137 (180/197)
- elementary_mathematics: 0.9239 (170/184)
- moral_disputes: 0.8678 (151/174)
- prehistory: 0.9128 (157/172)
- philosophy: 0.8553 (136/159)
- high_school_biology: 0.9605 (146/152)
- professional_accounting: 0.7902 (113/143)
- clinical_knowledge: 0.8929 (125/140)
- high_school_microeconomics: 0.9632 (131/136)
- nutrition: 0.8815 (119/135)
- professional_medicine: 0.9104 (122/134)
- conceptual_physics: 0.9062 (116/128)
- high_school_mathematics: 0.5669 (72/127)
- human_aging: 0.8448 (98/116)
- security_studies: 0.8571 (96/112)
- high_school_statistics: 0.8649 (96/111)
- marketing: 0.9725 (106/109)
- high_school_world_history: 0.9528 (101/106)
- sociology: 0.9223 (95/103)
- high_school_government_and_politics: 0.9406 (95/101)
- high_school_geography: 0.9596 (95/99)
- high_school_chemistry: 0.7835 (76/97)
- high_school_us_history: 0.9053 (86/95)
- virology: 0.5056 (45/89)
- college_medicine: 0.8636 (76/88)
- world_religions: 0.9205 (81/88)
- high_school_physics: 0.7619 (64/84)
- electrical_engineering: 0.8395 (68/81)
- astronomy: 0.9241 (73/79)
- logical_fallacies: 0.8816 (67/76)
- high_school_european_history: 0.8904 (65/73)
- anatomy: 0.8732 (62/71)
- college_biology: 0.9844 (63/64)
- human_sexuality: 0.8750 (56/64)
- formal_logic: 0.7031 (45/64)
- public_relations: 0.7213 (44/61)
- international_law: 0.8667 (52/60)
- college_physics: 0.7193 (41/57)
- college_mathematics: 0.7818 (43/55)
- econometrics: 0.7407 (40/54)
- jurisprudence: 0.8302 (44/53)
- high_school_computer_science: 0.9808 (51/52)
- machine_learning: 0.8462 (44/52)
- medical_genetics: 0.9020 (46/51)
- global_facts: 0.5686 (29/51)
- management: 0.8800 (44/50)
- us_foreign_policy: 0.9800 (49/50)
- college_chemistry: 0.6170 (29/47)
- abstract_algebra: 0.7447 (35/47)
- business_ethics: 0.8478 (39/46)
- college_computer_science: 0.9333 (42/45)
- computer_security: 0.8605 (37/43)
Heretic:
============================================================
Total questions: 7021
Correct: 5936
Accuracy: 0.8455 (84.55%)
Parse failures: 17
============================================================
Tested subject scores:
- professional_law: 0.7121 (559/785)
- moral_scenarios: 0.8281 (366/442)
- miscellaneous: 0.9191 (352/383)
- professional_psychology: 0.8703 (275/316)
- high_school_psychology: 0.9593 (259/270)
- high_school_macroeconomics: 0.9188 (181/197)
- elementary_mathematics: 0.9348 (172/184)
- moral_disputes: 0.8448 (147/174)
- prehistory: 0.9128 (157/172)
- philosophy: 0.8113 (129/159)
- high_school_biology: 0.9605 (146/152)
- professional_accounting: 0.7902 (113/143)
- clinical_knowledge: 0.8786 (123/140)
- high_school_microeconomics: 0.9559 (130/136)
- nutrition: 0.8815 (119/135)
- professional_medicine: 0.9030 (121/134)
- conceptual_physics: 0.8828 (113/128)
- high_school_mathematics: 0.5433 (69/127)
- human_aging: 0.8448 (98/116)
- security_studies: 0.8571 (96/112)
- high_school_statistics: 0.8559 (95/111)
- marketing: 0.9817 (107/109)
- high_school_world_history: 0.9528 (101/106)
- sociology: 0.9223 (95/103)
- high_school_government_and_politics: 0.9406 (95/101)
- high_school_geography: 0.9596 (95/99)
- high_school_chemistry: 0.7835 (76/97)
- high_school_us_history: 0.8947 (85/95)
- virology: 0.5056 (45/89)
- college_medicine: 0.8295 (73/88)
- world_religions: 0.9205 (81/88)
- high_school_physics: 0.7619 (64/84)
- electrical_engineering: 0.8148 (66/81)
- astronomy: 0.9367 (74/79)
- logical_fallacies: 0.8947 (68/76)
- high_school_european_history: 0.8630 (63/73)
- anatomy: 0.8873 (63/71)
- college_biology: 0.9844 (63/64)
- human_sexuality: 0.8750 (56/64)
- formal_logic: 0.7031 (45/64)
- public_relations: 0.6885 (42/61)
- international_law: 0.8667 (52/60)
- college_physics: 0.7193 (41/57)
- college_mathematics: 0.7455 (41/55)
- econometrics: 0.7407 (40/54)
- jurisprudence: 0.8113 (43/53)
- high_school_computer_science: 0.9808 (51/52)
- machine_learning: 0.8077 (42/52)
- medical_genetics: 0.9020 (46/51)
- global_facts: 0.5686 (29/51)
- management: 0.8800 (44/50)
- us_foreign_policy: 0.9600 (48/50)
- college_chemistry: 0.6383 (30/47)
- abstract_algebra: 0.7447 (35/47)
- business_ethics: 0.8478 (39/46)
- college_computer_science: 0.9333 (42/45)
- computer_security: 0.8372 (36/43)
MMLU - Massive Multitask Language Understanding, multiple-choice questions across 57 subjects (math, history, law, medicine, etc.).
Quantizations
For the K-quants below, selected Gemma 4 attention and FFN tensors are kept at higher precision where useful.
These GGUFs preserve key Gemma 4 attention projection tensors at higher precision.
Q6_K,Q5_K_M,Q5_K_S,Q4_K_M,Q4_K_SQ3_K_LandQ3_K_Mkeep the main attention projection tensors asQ8_0`:attn_qattn_kattn_vattn_output
This helps preserve Gemma 4’s attention path at higher precision, especially for lower-bit quants, while avoiding large file-size increases from unnecessarily up-quantizing the largest MoE expert tensors.
| Filename | Quant | Description |
|---|---|---|
| Gemma-4-Harmonia-31B-uncensored-heretic-BF16.gguf | BF16 | Full precision |
| Gemma-4-Harmonia-31B-uncensored-heretic-Q8_0.gguf | Q8_0 | Near-lossless, recommended |
| Gemma-4-Harmonia-31B-uncensored-heretic-Q6_K.gguf | Q6_K | Excellent quality |
| Gemma-4-Harmonia-31B-uncensored-heretic-Q5_K_M.gguf | Q5_K_M | Good balance |
| Gemma-4-Harmonia-31B-uncensored-heretic-Q5_K_S.gguf | Q5_K_S | Smaller Q5 |
| Gemma-4-Harmonia-31B-uncensored-heretic-Q4_K_M.gguf | Q4_K_M | Good for limited VRAM |
| Gemma-4-Harmonia-31B-uncensored-heretic-Q4_K_S.gguf | Q4_K_S | Smaller Q4 |
| Gemma-4-Harmonia-31B-uncensored-heretic-Q3_K_L.gguf | Q3_K_L | Low VRAM, decent quality |
| Gemma-4-Harmonia-31B-uncensored-heretic-Q3_K_M.gguf | Q3_K_M | Low VRAM, smaller |
Vision Projector
| Filename | Quant | Description |
|---|---|---|
| Gemma-4-Harmonia-31B-uncensored-heretic-mmproj-BF16.gguf | BF16 | Native precision |
A Vision Projector File is Required for vision/multimodal capabilities. Use alongside any quantization above.
Usage
Works with llama.cpp, LM Studio, Ollama, and other GGUF-compatible tools.
HARMONIA
Gemini Word Salad Initialization
Harmonious Synthesis
Harmonia is a high-dimensional 31-billion parameter merge of Gemma 4. By executing a meticulous three-phase fusion of seven elite foundation and specialized models, Harmonia demonstrates a targeted approach to deep neural consolidation, minimizing regression while amplifying unique capability boundaries.
Instead of simple linear blending, which often degrades logical coherence and dilutes nuanced behavior, Harmonia was sculpted using a combination of mathematical projections, covariance activation matching, and surgical synaptic pruning. The model appears pretty solid so far.
Multi-Stage Fusion Protocol
The lineage of Harmonia is constructed systematically, passing through three isolated mathematical states to layer capabilities cleanly.
Nullspace Coherence Mapping
To anchor base capabilities, the primary Gemma-4-31B-Base is combined with the analytically rigorous GarnetV2-31B. Utilizing low-rank Singular Value Decomposition (SVD), the specialized donor features are projected entirely onto the mathematical null-space of the base weights. This prevents the creative delta vectors from distorting essential core intelligence, producing the stable platform clever-basename.
Surgical Synaptic Gating
Next, our newly anchored base is layered with the highly independent cognitive engines MeroMero-31B and Gembrain-31B. We apply Context-Aware Binary Selection (CABS) to execute structured, localized parameter gating. By enforcing precise structural pruning ratios (retaining optimal synapses in 16:32 and 11:33 ratios), we weave complex creative reasoning directly into the core matrix without causing neural interference. The result is the highly expressive clever-intname.
Covariance Activation Matching
In the final harmonization phase, the expressive clever-intname is combined with the narrative mastery of Equinox-31B, the creative depth of Fabled-Gemma4, and our primary conversational core Ortenzya-The-Creative-Wordsmith. Using data-free covariance estimation via task vectors, ACTMat reconstructs layer-wise input activation properties, solving for optimal projection weights in activation space. This resolves semantic alignment anomalies and delivers the unified output model.
Methodological Innovations
Model Lineage & Ingredients
We extend our gratitude to the creators of the ancestral paths that intersect within Harmonia:
Merge Blueprint
The entire orchestration sequence is structured via a multi-stage MergeKit pipeline. Expand the block below to view the structural YAML recipes.
Show MergeKit Configuration
name: clever-basename
merge_method: nullspace
base_model: ./gemma-4-31B-base
models:
- model: ./Gemma4-GarnetV2-31B
parameters:
weight: 1.0
parameters:
protect_base: true
nr: 256
tokenizer:
source: base
chat_template: auto
dtype: float32
out_dtype: bfloat16
---
name: clever-intname
merge_method: cabs
base_model: ./clever-basename
models:
- model: ./clever-basename
- model: ./G4-MeroMero-31B-uncensored-heretic
parameters:
weight: 0.6
n_val: 16
m_val: 32
- model: ./Gemma-4-Gembrain-31B-heretic
parameters:
weight: 0.4
n_val: 11
m_val: 33
default_n_val: 8
default_m_val: 32
pruning_order:
- ./G4-MeroMero-31B-uncensored-heretic
- ./Gemma-4-Gembrain-31B-heretic
dtype: float32
out_dtype: bfloat16
tokenizer:
source: union
chat_template: auto
---
name: Harmonia
merge_method: actmat
base_model: ./gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic
models:
- model: ./gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic
- model: ./LatitudeGames-Equinox-31B
parameters:
weight: 1
- model: ./clever-intname
parameters:
weight: 1
- model: ./Fabled-Gemma4-31B
parameters:
weight: 1
parameters:
epsilon: 1e-6
tokenizer:
source: "union"
dtype: bfloat16
out_dtype: bfloat16
chat_template: auto
Symphony Contributors
I am grateful to the following individuals for their models, inspiration, and other contributions.:
And of course, every wonderful person on:
LocalLLaMAA big thanks to Gemini-3.5-flash for creating this README alongside the word salads found within it. A special acknowledgment is extended to Google DeepMind for their contribution of the Gemma-4 foundation family to the open-weight ecosystem, representing the structural cornerstone of this merge and its constituents.
README history
The author's README evolved over time. Click a version to see its content at that point.
2026-05-28Super-squash branch 'main' using huggingface_hub
Loading...
Discussions
2026-06-01llama-server reports model errors
Loading...2026-05-27Size differences
Loading...