G4-MeroMero-31B abliterated model lands on HuggingFace for local inference

G4-MeroMero-31B-uncensored-heretic-GGUF, a 31-billion-parameter abliterated language model from llmfan46, is now available in GGUF format for local inference. The checkpoint strips safety tuning and trains on zerofata's instruct-anime and gemini-3.1-pro-smallwiki datasets, positioning it for uncensored anime-adjacent generation and general instruction-following without content filters. GGUF quantization makes it compatible with llama.cpp, Koboldcpp, and other local inference engines that support the format.

At 31 billion parameters, the model sits in the sweet spot for consumer hardware—large enough for nuanced output, small enough to run on a single high-end GPU or across a pair of mid-tier cards with CPU offloading. Abliteration, a standard technique in the open-weight community, rewrites or removes specific attention heads responsible for refusal behavior, leaving underlying language capabilities intact. The instruct-anime dataset focuses on instruction-following in anime-style dialogue and scenario generation, while gemini-3.1-pro-smallwiki appears to be a distilled knowledge corpus. That pairing suggests dual-use tuning for both creative roleplay and factual question-answering, a common profile in the uncensored fine-tune scene. The model appeared on HuggingFace on May 15, 2025.