gemma-4-31B-it-uncensored-heretic-GGUF

llmfan46/gemma-4-31B-it-uncensored-heretic-GGUF provides GGUF quantizations (Q3_K_M through BF16) of a decensored variant of Google's Gemma 4 31B instruction-tuned model. The base weights were modified using Heretic v1.2.0 with Arbitrary-Rank Ablation (ARA), targeting attn.o_proj layers 30–48, achieving a 90% reduction in content refusals (10/100 vs 99/100 on the original) while maintaining a low KL divergence of 0.0541. Direct MMLU evaluation on 7,021 questions confirms 85.90% accuracy versus 86.50% for google/gemma-4-31B-it, demonstrating minimal capability loss from the decensoring process.

Built on Google's Gemma 4 architecture released April 2, 2026, this model inherits the full Gemma 4 31B capability set: 256K token context window, multimodal text-and-image input, native function calling, hybrid sliding-window/global attention, and support for 140+ languages. Official Google benchmarks for the base model include MMLU Pro 85.2%, LiveCodeBench v6 80.0%, AIME 2026 89.2%, GPQA Diamond 84.3%, and Codeforces ELO 2150. The base Gemma-4-31B-it ranks bronze (#3) on LMSYS Chatbot Arena with a score of 1449. HumanEval and GSM8K scores listed here are inherited estimates from independent community benchmarks of the base model (~85% each), adjusted for the measured 0.6% MMLU degradation.

The GGUF repository offers quantizations from Q3_K_M (~17GB) to BF16 (~62GB), plus separate mmproj files for vision inference. Recommended for local deployment via llama.cpp with --chat-template gemma. Suitable for users requiring reduced content filtering for creative writing, roleplay, or unrestricted drafting while retaining near-frontier open-model performance. Not listed on the official Open LLM Leaderboard; appears on OpenUGI uncensored leaderboard with a UGI score of 64.99 (with thought prefill). License: Apache 2.0.