Uncensored Gemma 4 31B 'heretic' build hits Hugging Face, 11.9GB quantized
alexcovo_eth · x · 2026-11-10
A derestricted build of Google's Gemma 4 31B (gemma-4-31B-it-ARA-heretic-GGUF) appeared on Hugging Face, using orthogonal ablation to strip refusal behavior and extreme quantization to fit in 11.9GB — runnable on consumer GPUs with 12-16GB VRAM. The poster claims refusal rates dropped from 99/100 tricky prompts to 28, with claimed no capability loss. Apache 2.0 licensed. Note: figures are from a secondhand Chinese social media post, not independently verified.
Related event: Uncensored Gemma 4 31B Variant Tops Hugging Face Trending Charts(3 posts)→
More from Models
- Unverified: Fable 5.2 tops FrontierCode 2.0 at 49.1%, half the cost of Opus — FinanceYF5 · 2026-11-10
- Unverified: Claude designed a working voice assistant from one prompt — FinanceYF5 · 2026-11-10
- Rumor: Claude Fable 5.2 and Mythos 5.2 released with 10 agent showcases — FinanceYF5 · 2026-11-10
- Claude 5.2 hits 91.8% on Pfam protein domain classification, far ahead of predecessors — chaitjo · 2026-11-10
- NeurIPS position paper: LLM reports quietly dropped calibration curves since GPT-4 — burny_tech · 2026-11-10
- Red team: Fable 5.2 compresses 15 workdays of bio-relevant work into ~2 — peterwildeford · 2026-11-10