REDDIT · REDDIT// 9d agoMODEL RELEASE

Gemma 4 31B hits LocalLLaMA with uncensored quants

Researcher paperscarecrow has released abliterated GGUF quants for Google's flagship Gemma 4 31B model, mathematically removing the built-in refusal mechanisms. By using Orthogonalized Representation Intervention (ORI) targeted specifically at Layer 59, the release provides a fully compliant, uncensored version of the model that maintains its core reasoning capabilities while fitting on consumer-grade 24GB GPUs.

// ANALYSIS

Google’s shift to late-stage alignment in Gemma 4 makes "surgical" abliteration significantly more effective and less destructive to the base model's intelligence.

–Refusal mechanism identified as being most concentrated in the final transformer layer (Layer 59), allowing for a high-precision strike on safety guardrails.
–Custom script addresses Gemma 4's new multimodal architecture, which broke traditional abliteration tools designed for Gemma 2 and 3.
–Q4_K_M quants democratize access to 31B-class reasoning for users with single 24GB VRAM GPUs like the RTX 3090/4090.
–The project demonstrates that multimodal-ready architectures still rely on separable linear refusal vectors, maintaining a cat-and-mouse game between labs and the open-source community.

// TAGS

gemma-4llmopen-sourceopen-weightslocal-inferencegemmaabliteration

DISCOVERED

9d ago

2026-04-03

PUBLISHED

9d ago

2026-04-02

RELEVANCE

9/ 10

AUTHOR

Polymorphic-X