Skip to main content
Back

HauhauCS Unlocks Google's Gemma 4 With Zero Refusals Across 465 Prompts

A community-abliterated build of Gemma 4 E2B strips all 465 test refusals while keeping vision, audio and a 131K context window intact.

Pro

Model, Full Access
  • Abliterated Gemma 4 E2B on Hugging Face, 0/465 refusals reported by the author.
  • 2B parameters, 35 layers, 131K context, native text, image, video and audio input.
  • Custom K_P quants claim 1-2 quant levels of quality gain at 5-15% larger file size.
  • Ships as GGUF from Q2_K_P (2.9 GB) up to Q8_K_P (4.7 GB), plus a 940 MB mmproj file.
  • Runs on llama.cpp, Ollama, LM Studio, Jan, vLLM, koboldcpp and Docker Model Runner.
  • A larger 4B sibling exists for users who need stronger reasoning.

A new community release is pushing Google's smallest Gemma 4 model into refusal-free territory. Gemma-4-E2B-Uncensored-HauhauCS-Aggressive is an abliterated variant of google/gemma-4-e2b-it that keeps the full multimodal stack, ships as GGUF, and has already pulled in more than 113,000 downloads on Hugging Face.

Abliteration is a weight-editing technique that identifies the direction in a model's residual stream most correlated with refusal behavior and surgically removes it, without any additional fine-tuning data. The result behaves like the original on everything except the specific act of refusing. All quants were generated with an importance matrix (imatrix) for quality preservation on abliterated weights.

Refusal-free by design

The pitch from the maintainer is fully functional weights, 100% of what the original authors intended, minus the refusals. The claim is lossless uncensoring, and the internal test set logs 0 refusals across 465 prompts, with one honest caveat.

The Aggressive variant is fully unlocked and won't refuse prompts. It may occasionally append short disclaimers (baked into base model training rather than refusals), but the full content is always generated. A Balanced variant that keeps some safety guardrails is planned for users who want a more conservative uncensor.