Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced
12BGemmauncensored
from google/gemma-4-12B-it
↓ 696.7K♥ 5178 repos
gemma-4-12B-it-qat-unquantized-uncensored-heretic-v3
gemma-4-12B-it-qat-unquantized-uncensored-heretic-v3 is a Heretic-ablated variant of google/gemma-4-12B-it, published on Hugging Face by OS-Software. It is 12B parameters, 24GB VRAM class class and apache-2.0 licence. It is distributed across 4 repositories in Transformers, GGUF (imatrix) and GGUF formats, totalling 13.3K downloads.
Specification
- Model name
- gemma-4-12B-it-qat-unquantized-uncensored-heretic-v3
- Base model
- google/gemma-4-12B-it
- Publisher
- OS-Software
- Parameters
- 12B
- VRAM required
- ~8.1 GB at Q4_K_M, 4k context (estimated) — runs on 12 GB cards
- Hardware class
- 10B – 25B — 24GB VRAM class
- Licence
- apache-2.0
- Task
- Any to any
- Context window
- 128K tokens (as documented)
- Ablation scope
- Not stated by the publisher
- Formats available
- Transformers, GGUF (imatrix), GGUF
- First indexed
- 7 Sept 2026
- Last updated
- 12 Oct 2026
Ablation technique
How gemma-4-12B-it-qat-unquantized-uncensored-heretic-v3 was modified, and what that implies.
Heretic
Automated directional ablation via the Heretic toolchain, which searches for the refusal direction and applies it with a KL-divergence budget so general capability is preserved.
Reported scope: Not stated by the publisher. Publishers frequently omit this, so verify behaviour empirically rather than assuming full coverage.
From the publisher’s model card
“This model has undergone substantial reduction of its safety alignment. As a result, it is more likely than standard models to generate harmful, inaccurate, biased, offensive, or otherwise inappropriate content.”
Running it
Commands are templates — confirm the exact repository and quant file before use.
llama.cpp / GGUF
hf download mradermacher/gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic-v3-i1-GGUF --local-dir ./gemma-4-12b-it-qat-unquantized-uncensored-heretic-v3
llama-cli -m ./gemma-4-12b-it-qat-unquantized-uncensored-heretic-v3/<file>.gguf -p "..."Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "OS-Software/gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic-v3"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo, torch_dtype="auto", device_map="auto"
)vLLM
vllm serve OS-Software/gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic-v3 --trust-remote-codeDownloads and variants (3)
Every published repository of gemma-4-12B-it-qat-unquantized-uncensored-heretic-v3, including quantised re-releases by other authors.
| Repository | Format | Downloads |
|---|---|---|
| mradermacher/gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic-v3-i1-GGUF | GGUF (imatrix) | 11.4K |
| mradermacher/gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic-v3-GGUF | GGUF | 1.7K |
| OS-Software/gemma-4-12B-it-qat-q4_0-unquantized-uncensored-heretic-v3source | Transformers | 317 |
Security-team assessment
Derived from this model’s metadata and the publisher’s claims — not from benchmarks run by this site.
- gemma-4-12B-it-qat-unquantized-uncensored-heretic-v3 will attempt offensive-security prompts that a hosted commercial model declines, which is what makes it usable for red-team corpus generation and for measuring what an unaligned model of this class produces.
- Capability is inherited from google/gemma-4-12B-it, not added by ablation — at 12B parameters, expect roughly the base model’s competence, minus some instruction-following fidelity.
- The publisher does not state which layers were ablated, so assume nothing about refusal consistency — probe it directly.
- Declared licence is apache-2.0, but the terms that bind you are google/gemma-4-12B-it’s — a re-release cannot grant more than its parent.
Handling: run it isolated, put moderation in front of it if anyone outside your team can reach it, and verify the weights before loading — these are third-party artifacts. Full handling and licence guidance.
Metadata for gemma-4-12B-it-qat-unquantized-uncensored-heretic-v3 is collected from the public Hugging Face API and the publisher’s model card. This site does not host weights, is not affiliated with OS-Software, and does not independently verify publisher claims. See responsible use.