Skip to main content
Back

huihui-ai Strips Google's Gemma-4-31B of All Safety Refusals

An uncensored, multimodal build of Google's Gemma-4-31B ships with both thinking and non-thinking modes stripped of refusals via weight editing.

  • Huihui-ai released an abliterated, uncensored build of Google's Gemma-4-31B-it under Apache 2.0.
  • Both thinking and non-thinking modes have refusals stripped via weight-level orthogonal projection, not fine-tuning.
  • Deployable via ollama run huihui_ai/gemma-4-abliterated:31B or Hugging Face Transformers multimodal loader.
  • Retains Gemma-4's image-text-to-text capability at 33B parameters in BF16, with GGUF quants available.
  • Recent research shows abliteration causes off-target effects like optimism bias and thinner uncertainty vocabulary.
  • Recommended for red-teaming, interpretability, and creative work, not production deployments requiring safety guarantees.

A new uncensored build of Google's Gemma-4-31B instruction-tuned model has landed on Hugging Face, and it is already pulling significant download traffic. The release, Huihui-gemma-4-31B-it-abliterated, is a 33B-parameter multimodal checkpoint that keeps Gemma-4's image-text-to-text capabilities intact while stripping out the model's built-in refusal behavior. It ships under Apache 2.0 in BF16 Safetensors, with a ready-to-pull Ollama build for anyone who wants to skip the Python plumbing.

Refusals, deleted at the weight level

Abliteration is a surgical weight edit rather than a fine-tune, grounded in a specific interpretability finding: refusal in aligned chat models is mediated by a single direction in the residual stream, consistent across many open-source models up to 72B parameters. Ablating that direction at every layer and token position via orthogonal projection prevents the model from refusing harmful instructions while preserving general capabilities. The result is a permanent weight edit that requires no retraining.

The recipe is straightforward in principle:

  1. Compute the mean difference between hidden states produced by harmful and harmless prompts to obtain a refusal direction.
  2. Project that direction out of the model's weight matrices.