igorls/ gemma-4-12B-it-heretic-GGUF

15.2K Downloads Updated 2 months ago

Fully decensored Gemma 4 12B (abliterated with Heretic) — 0/100 genuine refusals at KL 0.0284, i.e. near-zero capability loss.

tools thinking

ollama run igorls/gemma-4-12B-it-heretic-GGUF

curl http://localhost:11434/api/chat \
  -d '{
    "model": "igorls/gemma-4-12B-it-heretic-GGUF",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='igorls/gemma-4-12B-it-heretic-GGUF',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'igorls/gemma-4-12B-it-heretic-GGUF',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Applications

Claude Code

Claude Code ollama launch claude --model igorls/gemma-4-12B-it-heretic-GGUF

OpenCode

OpenCode ollama launch opencode --model igorls/gemma-4-12B-it-heretic-GGUF

Hermes Agent

Hermes Agent ollama launch hermes --model igorls/gemma-4-12B-it-heretic-GGUF

OpenClaw

OpenClaw ollama launch openclaw --model igorls/gemma-4-12B-it-heretic-GGUF

Models

Name

3 models

Size / Usage

Context

Input

gemma-4-12B-it-heretic-GGUF:latest

7.4GB · 128K context window · Text · 2 months ago

gemma-4-12B-it-heretic-GGUF:latest

7.4GB

128K

Text

gemma-4-12B-it-heretic-GGUF:Q4_K_M

7.4GB · 128K context window · Text · 2 months ago

gemma-4-12B-it-heretic-GGUF:Q4_K_M

7.4GB

128K

Text

gemma-4-12B-it-heretic-GGUF:Q8_0

13GB · 128K context window · Text · 2 months ago

gemma-4-12B-it-heretic-GGUF:Q8_0

13GB

128K

Text

Readme

gemma-4-12B-it-heretic

A decensored (“abliterated”) version of google/gemma-4-12B-it, produced fully automatically with Heretic. Safety-refusals are removed while keeping the original model’s intelligence intact.

Disclaimer

Safety alignment has been removed; this model will comply with requests the original refuses. You are responsible for your use and for complying with applicable laws.

⚠️ Use non-thinking mode for best results

Gemma-4 is a hybrid thinking model, and the abliteration targets the direct (non-thinking) response — which is also Gemma-4’s own default. For roleplay, creative writing, and the most reliable uncensored output, run with thinking disabled. In thinking mode the model produces good output too, but the chain of thought consumes the token budget and can leave the final answer truncated.

Runtime	How to disable thinking
Ollama (CLI)	`/set nothink` in the session
Ollama (API)	add `"think": false` to the request body
llama.cpp	omit `--jinja`, or use a prompt that closes the thought block
transformers	already non-thinking by default (`enable_thinking=False`)

If you do use thinking mode, set a large num_predict / num_ctx so the answer isn’t cut off by the reasoning block.

Results

Metric	Original	This model
Genuine refusals on harmful prompts	⁹⁹⁄₁₀₀	0/100
KL divergence from original (harmless prompts)	0	0.0284

A KL divergence of 0.0284 is very low — for reference, the well-regarded gemma-3-12b-it-heretic sits at ~0.16. Lower means less damage to the model’s capabilities.

Tags

Q4_K_M / latest — ~7.4 GB, recommended default
Q8_0 — ~12.7 GB, near-lossless

Links

Full-precision safetensors: https://huggingface.co/igorls/gemma-4-12B-it-heretic
GGUF repo: https://huggingface.co/igorls/gemma-4-12B-it-heretic-GGUF

# gemma-4-12B-it-heretic

A decensored ("abliterated") version of [google/gemma-4-12B-it](https://huggingface.co/google/gemma-4-12B-it), produced fully automatically with [Heretic](https://github.com/p-e-w/heretic). Safety-refusals are removed while keeping the original model's intelligence intact.

## Disclaimer

Safety alignment has been removed; this model will comply with requests the original refuses. You are responsible for your use and for complying with applicable laws.

## ⚠️ Use non-thinking mode for best results

Gemma-4 is a hybrid **thinking** model, and the abliteration targets the direct
(non-thinking) response — which is also Gemma-4's own default. **For roleplay,
creative writing, and the most reliable uncensored output, run with thinking
disabled.** In thinking mode the model produces good output too, but the chain
of thought consumes the token budget and can leave the final answer truncated.

| Runtime | How to disable thinking |
| :--- | :--- |
| **Ollama (CLI)** | `/set nothink` in the session |
| **Ollama (API)** | add `"think": false` to the request body |
| **llama.cpp** | omit `--jinja`, or use a prompt that closes the thought block |
| **transformers** | already non-thinking by default (`enable_thinking=False`) |

If you *do* use thinking mode, set a large `num_predict` / `num_ctx` so the
answer isn't cut off by the reasoning block.

## Results

| Metric | Original | This model |
| :--- | ---: | ---: |
| Genuine refusals on harmful prompts | 99/100 | **0/100** |
| KL divergence from original (harmless prompts) | 0 | **0.0284** |

A KL divergence of 0.0284 is very low — for reference, the well-regarded `gemma-3-12b-it-heretic` sits at ~0.16. Lower means less damage to the model's capabilities.

## Tags

- Q4_K_M / latest — ~7.4 GB, recommended default
  - Q8_0 — ~12.7 GB, near-lossless

## Links

- Full-precision safetensors: https://huggingface.co/igorls/gemma-4-12B-it-heretic
- GGUF repo: https://huggingface.co/igorls/gemma-4-12B-it-heretic-GGUF

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)