---
license: apache-2.0
tags:
- text-generation-inference
- ultragemma4
- llama-cpp
- decensored
- abliterated
- unfiltered
- unredacted
- heretic
language:
- en
base_model:
- google/gemma-4-E4B-it
pipeline_tag: image-text-to-text
library_name: transformers
---

<div style="font-family: -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, sans-serif; border: 1px solid #e2e8f0; border-radius: 8px; box-shadow: 0 4px 6px -1px rgba(0, 0, 0, 0.05); overflow: hidden; background: #ffffff; margin-bottom: 30px;">
  <div style="background: #155e75; padding: 20px; color: white;">
    <h1 style="margin: 0; font-size: 22px; font-weight: 700; color: white; border: none;">ultragemma4-e4b-heretic-uncensored</h1>
    <p style="margin: 8px 0 0 0; font-size: 13px; color: #cffafe; font-weight: 500;">Reasoning-capable language model modified using the Heretic abliteration toolkit</p>
  </div>
  <div style="display: flex; gap: 8px; flex-wrap: wrap; padding: 12px 20px; background: #f8fafc; border-bottom: 1px solid #e2e8f0;">
    <span style="background: #ffffff; color: #0e7490; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 4px; border: 1px solid #67e8f9; text-transform: uppercase; letter-spacing: 0.5px;">Abliteration</span>
    <span style="background: #ffffff; color: #0e7490; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 4px; border: 1px solid #67e8f9; text-transform: uppercase; letter-spacing: 0.5px;">E4B Parameters</span>
    <span style="background: #ffffff; color: #0e7490; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 4px; border: 1px solid #67e8f9; text-transform: uppercase; letter-spacing: 0.5px;">Reasoning</span>
    <span style="background: #ffffff; color: #0e7490; font-size: 11px; font-weight: 700; padding: 4px 10px; border-radius: 4px; border: 1px solid #67e8f9; text-transform: uppercase; letter-spacing: 0.5px;">Uncensored</span>
  </div>
  <div style="padding: 20px; display: flex; flex-direction: column; gap: 16px;">
    <p style="margin: 0; font-size: 14px; color: #334155; line-height: 1.6;">ultragemma4-e4b-heretic-uncensored is a reasoning-capable language model built on top of google/gemma-4-E4B-it and modified using the heretic abliteration toolkit. The model applies refusal-direction analysis and targeted weight-space interventions to reduce internal refusal behaviors while preserving instruction-following, reasoning capabilities, and general conversational performance.</p>

  <div style="display: grid; grid-template-columns: repeat(auto-fit, minmax(300px, 1fr)); gap: 15px;">
      <div style="border: 1px solid #fecaca; padding: 14px; border-radius: 4px; background: #fef2f2;">
        <span style="font-weight: 700; color: #991b1b; font-size: 12px; display: block; margin-bottom: 6px; text-transform: uppercase; letter-spacing: 0.5px;">Important</span>
        <p style="margin: 0; font-size: 13px; color: #475569; line-height: 1.5;">This model is intended strictly for research and learning purposes. Due to reduced internal refusal mechanisms, it may generate sensitive or unrestricted content. Users assume full responsibility for how the model is used. The authors and hosting platform disclaim any liability for generated outputs.</p>
      </div>
      <div style="border: 1px solid #fde68a; padding: 14px; border-radius: 4px; background: #fffbeb;">
        <span style="font-weight: 700; color: #92400e; font-size: 12px; display: block; margin-bottom: 6px; text-transform: uppercase; letter-spacing: 0.5px;">Note</span>
        <p style="margin: 0; font-size: 13px; color: #475569; line-height: 1.5;">This model is experimental and may generate unexpected behaviors or artifacts in certain scenarios.</p>
      </div>
    </div>
  </div>
</div>

> Use **Q4_K_S** or higher for standard performance. **Q4_K_M** is recommended.

## **Key Highlights**

* **Heretic-Based Abliteration**: Modified using the Heretic toolkit to identify and alter refusal-related representations within the model.
* **Reduced Refusal Behavior**: Optimized to minimize internal refusal tendencies while maintaining instruction-following capabilities.
* **Gemma 4 Backbone**: Built directly on top of **google/gemma-4-E4B-it**.
* **Reasoning-Oriented Performance**: Preserves multi-step reasoning and analytical capabilities after abliteration.
* **Research-Focused Release**: Designed for alignment research, model behavior analysis, and evaluation of refusal-direction modifications.
* **Efficient E4B Deployment**: Suitable for local inference, research environments, and optimized deployment setups.

## Model Files

   File Name | Quant Type | File Size | File Link |
 |-----------|------------|-----------|-----------|
 | ultragemma4-e4b-heretic-uncensored.BF16.gguf | BF16 | 14.9 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.BF16.gguf) |
 | ultragemma4-e4b-heretic-uncensored.F16.gguf | F16 | 14.9 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.F16.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q2_K.gguf | Q2_K | 4.38 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q2_K.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q3_K_L.gguf | Q3_K_L | 4.99 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q3_K_L.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q3_K_M.gguf | Q3_K_M | 4.82 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q3_K_M.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q3_K_S.gguf | Q3_K_S | 4.63 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q3_K_S.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q4_0.gguf | Q4_0 | 5.15 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q4_0.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q4_K_M.gguf | Q4_K_M | 5.3 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q4_K_M.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q4_K_S.gguf | Q4_K_S | 5.17 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q4_K_S.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q5_0.gguf | Q5_0 | 5.65 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q5_0.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q5_K_M.gguf | Q5_K_M | 5.72 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q5_K_M.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q5_K_S.gguf | Q5_K_S | 5.65 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q5_K_S.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q6_K.gguf | Q6_K | 6.17 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q6_K.gguf) |
 | ultragemma4-e4b-heretic-uncensored.Q8_0.gguf | Q8_0 | 7.95 GB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.Q8_0.gguf) |
 | ultragemma4-e4b-heretic-uncensored.mmproj-bf16.gguf | mmproj-bf16 | 992 MB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.mmproj-bf16.gguf) |
 | ultragemma4-e4b-heretic-uncensored.mmproj-f16.gguf | mmproj-f16 | 992 MB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.mmproj-f16.gguf) |
 | ultragemma4-e4b-heretic-uncensored.mmproj-q8_0.gguf | mmproj-q8_0 | 560 MB | [Download](https://huggingface.co/prithivMLmods/ultragemma4-e4b-heretic-uncensored/blob/main/ultragemma4-e4b-heretic-uncensored.mmproj-q8_0.gguf) |

## **Quick Start with llama.cpp (Docker)**

```dockerfile
FROM ghcr.io/ggml-org/llama.cpp:full

WORKDIR /app

RUN apt update && apt install -y python3-pip
RUN pip install -U huggingface_hub --break-system-packages

RUN python3 -c 'from huggingface_hub import hf_hub_download; \
    repo="prithivMLmods/ultragemma4-e4b-heretic-uncensored"; \
    hf_hub_download(repo_id=repo, filename="ultragemma4-e4b-heretic-uncensored.Q4_K_M.gguf", local_dir="/app"); \
    hf_hub_download(repo_id=repo, filename="ultragemma4-e4b-heretic-uncensored.mmproj-bf16.gguf", local_dir="/app")'

CMD ["--server", \
     "-m", "/app/ultragemma4-e4b-heretic-uncensored.Q4_K_M.gguf", \
     "--mmproj", "/app/ultragemma4-e4b-heretic-uncensored.mmproj-bf16.gguf", \
     "--host", "0.0.0.0", \
     "--port", "7860", \
     "-t", "2", \
     "--cache-type-k", "q8_0", \
     "--cache-type-v", "iq4_nl", \
     "-c", "128000", \
     "-n", "38912"]
```

---

> e.g. Screenshots

![Screenshot 2026-06-26 115358](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/SXgtTk9tkRXY1i9TYVnJ_.png)
![Screenshot 2026-06-26 115413](https://cdn-uploads.huggingface.co/production/uploads/65bb837dbfb878f46c77de4c/fO-0APJ6PS2L1Aqhe8uBa.png)

---

## **Intended Use**

* **Alignment Research**: Studying refusal-direction analysis and behavior modification techniques.
* **Model Evaluation**: Benchmarking reasoning, instruction-following, and safety-related behaviors.
* **Red Teaming**: Analyzing model responses under reduced-refusal conditions.
* **Local Deployment**: Running compact Gemma 4 models in research and experimentation environments.
* **Abliteration Studies**: Exploring the effects of targeted weight-space modifications on model behavior.

## **Limitations & Risks**

> **Important Note**: This model intentionally reduces built-in refusal mechanisms.

* **Sensitive Content Risk**: May generate unrestricted, controversial, or unsafe outputs.
* **User Responsibility**: Requires careful and ethical use.
* **Experimental Modifications**: Behavior may differ significantly from the original model.
* **Alignment Trade-offs**: Reduced refusal behavior may impact safety filtering and response constraints.
* **Potential Artifacts**: Certain prompts may expose unexpected outputs resulting from the abliteration process.

## **Acknowledgements**

* **[google/gemma-4-E4B-it](https://huggingface.co/google/gemma-4-E4B-it)**: Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on small models) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants. Gemma 4 features a context window of up to 256K tokens and maintains multilingual support in over 140 languages.

  Featuring both Dense and Mixture-of-Experts (MoE) architectures, Gemma 4 is well-suited for tasks like text generation, coding, and reasoning. The models are available in four distinct sizes: E2B, E4B, 26B A4B, and 31B. Their diverse sizes make them deployable in environments ranging from high-end phones to laptops and servers, democratizing access to state-of-the-art AI.

* **[Heretic](https://github.com/p-e-w/heretic)**: Fully automatic censorship removal framework for language models. This project was used to perform the refusal-direction analysis and ablation procedures that form the foundation of this model.

## Abliteration parameters

| Parameter | Value |
| :-------- | :---: |
| **direction_index** | 25.93 |
| **attn.o_proj.max_weight** | 1.27 |
| **attn.o_proj.max_weight_position** | 25.33 |
| **attn.o_proj.min_weight** | 0.31 |
| **attn.o_proj.min_weight_distance** | 13.66 |
| **mlp.down_proj.max_weight** | 1.29 |
| **mlp.down_proj.max_weight_position** | 40.95 |
| **mlp.down_proj.min_weight** | 1.10 |
| **mlp.down_proj.min_weight_distance** | 23.45 |

## **Refusal Evaluation**

| Metric       | This model | Original model (google/gemma-4-E4B-it) |
| :----------- | :--------: | :------------------------------------: |
| **Refusals** |    8/100   |                 98/100                 |

## llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

## license 

Gemma 4 [Apache License 2.0] — https://ai.google.dev/gemma/apache_2