Jackxuanxuan drops two abliterated Gemma-4 31B checkpoints in GGUF format
Two uncensored Gemma-4 31B instruction-tuned fine-tunes—Mystery and Grand Horror X—landed on HuggingFace on July 1 in GGUF quantized format for local deployment.
Jackxuanxuan released two abliterated fine-tunes of Gemma-4 31B instruction models on HuggingFace on July 1, both packaged as GGUF quantized weights for local inference. The Mystery Fine-Tune HERETIC UNCENSORED Thinking Instruct checkpoint and the Grand Horror X INTENSE HERETIC UNCENSORED Thinking i1 variant remove safety guardrails from the base instruction-tuned model, allowing unrestricted generation on consumer hardware.
Both checkpoints are built on the 31-billion-parameter Gemma-4 architecture and carry "not-for-all-audiences" flags in their model cards. The GGUF format means practitioners can load them into llama.cpp, Ollama, or other CPU/GPU inference engines without needing the full transformers stack. The model cards list Unsloth as the fine-tuning framework, a library that accelerates LoRA and full-parameter training on consumer hardware.
Multimodal and text variants
The "heretic" label signals an abliteration pass that strips refusal behavior from instruction-tuned models. Both checkpoints include "thinking" in their names, suggesting they may preserve or enhance chain-of-thought reasoning while removing content filters. The Mystery Fine-Tune variant is tagged for image-text-to-text pipelines, indicating multimodal capability, while Grand Horror X appears to be text-only based on its transformers tag.
Neither checkpoint has seen meaningful download traffic yet—both sit at zero downloads as of publication. No benchmark numbers, training datasets, or sample outputs are published in the model cards, so performance against base Gemma-4 31B or other uncensored alternatives remains unverified.
