Uncensored Gemma 4 31B hits Hugging Face: refusals drop from 99/100 to 28
udmrzn · x · 2026-11-10
A community-built uncensored variant of Google's official Gemma 4 31B, "gemma-4-31B-it-ARA-heretic" (GGUF quants by mradermacher), is now on Hugging Face under Apache 2.0, compressed to just 11.9GB so it can run locally on consumer PCs with 12-16GB VRAM.
Key points:
- Refusal rate collapsed: on 100 difficult prompts, the original refused 99; the abliterated version refused only 28
- Method: orthogonal ablation (ARA) surgically removes refusal neurons, cleaning moralizing from both direct-answer and deep-thinking modes with reportedly no capability loss
- Use cases: pentest script writing, sensitive source analysis, unrestricted long-form writing
- Availability: free download, no license gates
More from Models
- Claude Fable 5.2 and Mythos 5.2 spotted on AWS Bedrock, priced 20% below 5.1 — testingcatalog · 2026-11-10
- Meta's Muse Ember hits efficiency frontier: Haiku-level 36 score at 1/3 the cost — giffmana · 2026-11-10
- Meta's Muse Ember finds $2.48M misallocation across 13 fund documents in Box — alexandr_wang · 2026-11-10
- MiniMax H3 video generation running locally on just 6GB VRAM — No_Cow26 · 2026-11-10
- If Meta squeezed that much intelligence per compute into a small model, larger ones should follow — daniel_mac8 · 2026-11-10
- Ripple-65B-Preview tops open-weight agentic board with 98.2% job completion — ryanshrout · 2026-11-10