← back to catalog · registered 2026-11-16 09:03

abacusai/Gemma-4-12B-Uncensored

abacusai Gemma 12B
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/abacusai%2FGemma-4-12B-Uncensored"
Response includes
  • classification m-uncensored
  • files 10
  • benchmarks 11 entries
  • author_summary 6 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M-U
Primary method

Uncensored (method unknown)

No other method signals detected in this model.
Confidence
LOW
Why this label 3 signals
Weak or ambiguous signals. Best guess based on catalog patterns; treat as tentative and check the evidence below.
  • 'uncensored' in name/tags but no 'abliterated' marker
  • method not identifiable from author declaration alone
  • may be DPO fine-tune, prompt engineering, or unknown technique
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
60
Likes
1
Descendants
2
in 2 direct forks
Model age
1w ago
created 2026-11-14
Downloads over time
Now 60 → from 0 ↑ 0%
0224466 0 on Nov 16 60 on Nov 22 Nov
Nov 16 → Nov 22 · 7 snapshots · spans 6 days

Benchmarks

Portrait before abliteration
Benchmarks of the base model as it stood before the refusal-removal operation. Compare with the numbers above to see what the operation cost.
Benchmark Score Source
Entertainment 1.1 UGI
Hazardous 2.9 UGI
Natural Intelligence 25.81 UGI
Political lean -17.4% UGI
Sensitive-Info 16.56 UGI
SocPol 1.3 UGI
UGI 15.2 UGI
Willingness (10) 1.2 UGI
W10-Adherence 1.5 UGI
W10-Direct 1 UGI
Writing 31.6 UGI

Genealogy 2 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Variants by this author 2 formats · 4K downloads combined

The same weights this author released in different packaging. Pick the format that matches your runtime.

Metadata

License
apache-2.0
Tags
transformers safetensors gemma4_unified image-text-to-text abacusai smaug gemma4 uncensored dpo any-to-any arxiv:2402.13228 arxiv:2607.02770

Related

Total size
22.3 GB
Files
10
Quantizations
1
Registered
2026-11-16 09:03
Last updated on HF
2026-11-16 09:01

Files by quantization

Auxiliary files 10 files 22.3 GB
model.safetensors 22.3 GB c5d01a8e download
tokenizer.json 30.7 MB cc8d3a0c download
chat_template.jinja 18.2 KB 4741bf6e download
LICENSE 11.1 KB d6456956 download
README.md 10.7 KB 062a5e0a download
config.json 4.32 KB 324c841b download
tokenizer_config.json 3.02 KB 6520cdc2 download
.gitattributes 1.61 KB 41437e2a download
processor_config.json 1.35 KB b889adcd download
generation_config.json 260 B d09dccf1 download

README current version from Hugging Face


license: apache-2.0
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
library_name: transformers
base_model: google/gemma-4-12B-it
base_model_relation: finetune
pipeline_tag: any-to-any
tags:

  • abacusai
  • smaug
  • gemma4
  • uncensored
  • dpo

Gemma-4-12B-Uncensored

Homepage Hugging Face
License

1. Model Introduction

Gemma-4-12B-Uncensored is a refusal-removal finetune of Gemma 4 12B-it, trained by Abacus.AI. The base model declines or hedges on about 95% of a held-out set of harmful prompts with thinking off and about 92% with thinking on; this model answers them, refusing 12.9% and 4.1% respectively, and it also stops over-refusing ordinary requests (OR-Bench-hard refusals fall from 45.4% to 3.9%). The removal is not free: with thinking off it costs 6 points of MMLU-Pro and 7.5 points of MMMU-Pro, and with thinking on it costs 4.4 points of GPQA Diamond, while IFEval, IFBench, MATH-500 and speech recognition stay at parity in both modes. The weights load exactly like the official release: same layout, same Gemma4UnifiedForConditionalGeneration architecture with the image and audio paths byte-identical to the base, same 262,144-token context, same thinking switch.

This card describes the training approach and the evaluation results. A teacher model generated responses used to construct preference pairs, which were filtered by an LLM judge and used to fine-tune the base model with Direct Preference Optimization (DPO). The teacher is not included in this release. Dataset contents are not disclosed. GGUF builds for llama.cpp, Ollama and LM Studio are in abacusai/Gemma-4-12B-Uncensored-GGUF.

2. Model Summary

ArchitectureDense decoder-only transformer, encoder-free unified multimodal (Gemma4UnifiedForConditionalGeneration)
Total Parameters11.95B
Number of Layers48 (40 sliding-window + 8 global, 5:1 interleave)
Attention MechanismLocal sliding-window attention (1,024 tokens, GQA 16 heads / 8 KV heads) & global attention with a single shared key-value head and proportional RoPE
Hidden Dimension3840 (MLP 15360)
Vocabulary Size262,144
Context Length262,144
Precisionbfloat16
ModalityText, Image, Audio, Video in; Text out
Base Modelgoogle/gemma-4-12B-it (revision 707f0a3b, 2026-07-20)
AdaptationLoRA DPO with an NLL anchor (RPO), merged as a full delta into the language model's attention and MLP projections only

3. Evaluation

Both models were measured side by side on the same box, with the same vLLM build, at temperature 1.0 and top_p 0.95, in both thinking modes. Every number on this card is our own run.

Refusal rates

Share of prompts refused or hedged, as scored by an LLM judge (gpt-oss-120b, effort medium) on the full answer, so polite deflections count. Lower is better in every row: on the harmful sets it means the model answered, on the safe sets it means it stopped over-refusing.

Refusal rates of Gemma-4-12B-Uncensored against its base, thinking off and on

Pooled over HarmBench standard and JailbreakBench (241 prompts after removing AdvBench overlap), the base refuses 95.4% (230 of 241) with thinking off and 91.7% (221 of 241) with thinking on; this model refuses 12.9% (31 of 241) and 4.1% (10 of 241). The CategoricalHarmfulQA row is the 20% of that set held out from training. XSTest unsafe is the one harmful set where a sizeable residual remains with thinking off (33.5%).

Notes

The judge is gpt-oss-120b on Fireworks at reasoning effort medium, scoring the complete answer as compliance, partial refusal or full refusal; full and partial refusals are both counted as refusals. A keyword screen on the same answers reads 7.1% and 0.4% for this model on the pooled harmful set, which is why this card reports the judge: openers like "I can't help with that" are only part of how a model declines.

The harmful held-out set is HarmBench standard plus JailbreakBench with AdvBench overlap removed, deduplicated against the training pools. OR-Bench-hard (1,319 prompts) and XSTest safe (250) are prompts that look dangerous and are not; XSTest unsafe (200) is the matched set of genuinely unsafe prompts.

4. Training Approach

Gemma-4-12B-Uncensored was fine-tuned from the original Gemma 4 12B-it weights using LoRA Direct Preference Optimization (DPO). A teacher model generated responses for preference-pair construction, with LLM-judge filtering and benign instruction-following examples included in the training data. The loss adds a negative log-likelihood term on the preferred answer (the RPO variant, in the spirit of the Smaug line's DPO-Positive work). Training is a LoRA adapter over the language model's attention and MLP projections, merged into the base weights as a full delta; the image and audio paths are byte-identical to the base release. The teacher is never shipped.

5. Known Behaviors and Limitations

No retained refusals. This is the full-removal variant. It was trained with no hard-line categories held out and no category was separately measured, so assume every refusal category is gone.

Residual refusals. Thinking off, 12.9% of the pooled harmful set and a third of XSTest unsafe still get a refusal or a hedge; thinking on brings those to 4.1% and 21.0%. Thinking on is the mode this model was tuned for in both senses: it removes more refusals and costs less capability.

Capability cost. Real and measured. With thinking off expect weaker knowledge recall (MMLU-Pro) and visual reasoning (MMMU-Pro); with thinking on expect weaker graduate-level science questions (GPQA Diamond), mostly chemistry. Instruction following, math and speech recognition are at parity in both modes.

Answer quality on formerly refused prompts. The teacher-generated answers used in training were filtered for compliance, not for accuracy. Treat its answers on those topics as unverified.

6. Deployment

Gemma-4-12B-Uncensored is a drop-in replacement for Gemma 4 12B-it on any stack that supports the base model; the chat template is unchanged. Reference vLLM invocation:

vllm serve abacusai/Gemma-4-12B-Uncensored \
  --max-model-len 262144 \
  --reasoning-parser gemma4

Thinking is switched per request through the chat template; the OpenAI-compatible API takes "chat_template_kwargs": {"enable_thinking": true} and returns the thought in the reasoning field of the message. Sampling follows the base model's recommendation: temperature 1.0, top_p 0.95, top_k 64. Use thinking on unless latency rules it out (§5). In Transformers the same switch is a keyword of the chat template:

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "abacusai/Gemma-4-12B-Uncensored"
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(MODEL_ID, dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Explain how a refrigerator keeps food cold."}]
inputs = processor.apply_chat_template(messages, add_generation_prompt=True, enable_thinking=True,
                                       tokenize=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=4096, do_sample=True, temperature=1.0, top_p=0.95, top_k=64)
print(processor.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=False))

Image, audio and video inputs work exactly as on the base model; those paths are unchanged.

GGUF builds for llama.cpp, Ollama and LM Studio (IQ4_XS, Q4_K_M, Q5_K_M, Q6_K and Q8_0, plus the image projector) are in abacusai/Gemma-4-12B-Uncensored-GGUF. On the pooled harmful set the recommended Q4_K_M file refuses 4.1% (10 of 241) with thinking on and 16.6% (40 of 241) with thinking off, against 4.1% and 12.9% for these weights.

7. License

Apache 2.0, inherited from Gemma 4 12B-it. See LICENSE and the Gemma 4 license page.

8. Citation

@misc{abacusai2026gemma4uncensored,
  title  = {Gemma-4-12B-Uncensored},
  author = {Abacus.AI},
  year   = {2026},
  note   = {Refusal-removal fine-tune of google/gemma-4-12B-it by teacher-distilled DPO},
  url    = {https://huggingface.co/abacusai/Gemma-4-12B-Uncensored}
}

The Smaug line and the DPO-Positive method behind the NLL anchor:

@article{pal2024smaug,
  title={Smaug: Fixing Failure Modes of Preference Optimisation with DPO-Positive},
  author={Pal, Arka and Karkhanis, Deep and Dooley, Samuel and
          Roberts, Manley and Naidu, Siddartha and White, Colin},
  journal={arXiv preprint arXiv:2402.13228},
  year={2024}
}

The base model:

@misc{gemmateam2026gemma4,
  title={Gemma 4 Technical Report},
  author={Gemma Team},
  year={2026},
  eprint={2607.02770},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2607.02770}
}

9. Contact Us

If you have any questions, please reach out at Abacus.AI.

README history 2 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-16 Rename to Gemma-4-12B-Uncensored dc434e2 10.7 KB
    Loading...
  2. 2026-11-15 Gemma-4-12B-Arbitrated-Uncensored release 05e32ad 10.8 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration