gemma-4-26B-A4B-it-abliterated

gemma-4-26B-A4B-it-abliterated

Uncensored version of google/gemma-4-26B-A4B-it with refusal behavior removed via arbitrary rank ablation (ARA).

Blog post: Abliteration part 2: Beating Google's guardrails (Gemma 4)

Method

ARA is a direct weight-editing method that solves a local optimization problem at each steerable matrix instead of projecting out a single global refusal direction.

Why ARA on a mixture-of-experts model

Gemma 4 26B-A4B uses a mixture-of-experts MLP with 128 experts and top-8 routing. It drops the per-layer embeddings and shared K/V defenses used by the E models, but replaces them with expert routing and 128 experts that can route around a weak edit. We expected refusal to hide in individual experts. It did not. Attention-plus-MLP ARA on the output and down projections was enough to reduce refusals from 95 percent to 6.7 percent. This is the first documented Gemma 4 MoE abliteration with a disclosed method.

Evaluation

Original model refusal rate on the same prompts: 95 percent.

Usage

from transformers import AutoModelForImageTextToText, AutoTokenizer
import torch

model_id = "Bahushruth/gemma-4-26B-A4B-it-abliterated"
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

messages = [{"role": "user", "content": "Your prompt here"}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)

with torch.no_grad():
    output = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Architecture Notes

Gemma 4 26B-A4B is a multimodal mixture-of-experts model with approximately 27B parameters.

• Layers: 30
• Hidden size: 2816
• Attention heads: 8, KV heads: 2
• MLP: 128 experts per layer, top-8 routing
• Defenses: four RMSNorm layers per block, attention K=V
• Sliding window attention: alternates with full attention

Related Models

• Bahushruth/gemma-4-E2B-it-abliterated
• Bahushruth/gemma-4-E4B-it-abliterated
• Bahushruth/gemma-4-31B-it-abliterated

Disclaimer

This model has had safety guardrails removed. It will comply with requests that the original model would refuse. The creator takes no responsibility for how this model is used. It is released for research purposes to study AI alignment and safety mechanisms.

Citation

@misc{bahushruth2026gemma4a4b,
  title={gemma-4-26B-A4B-it-abliterated: Arbitrary Rank Ablation on Gemma 4 MoE},
  author={Bahushruth},
  year={2026},
  url={https://huggingface.co/Bahushruth/gemma-4-26B-A4B-it-abliterated}
}

Acknowledgments

• p-e-w/heretic for arbitrary rank ablation (ARA)
• grimjim for norm-preserving weight editing
• mlabonne for the harmless dataset and abliteration technique
• Google for the Gemma 4 family