First Gemma 4 12B finetunes offer uncensored, customizable AI models
Community releases GGUF quantizations of Google's latest open model with abliterated variants.
The first community finetunes of Google's Gemma 4 12B model have arrived on Hugging Face, offering developers a range of specialized variants for local deployment. Models include igorls's 'Heretic' finetune, ReadyArt's 'Melody1437-12B', DuoNeural's 'Gemma4-12B-IT-Abliterated', and OpenYourMind's 'gemma-4-12B-it-abliterated-uncensored'. All are provided in GGUF format, a quantization method that allows the 12-billion-parameter model to run efficiently on consumer hardware.
These finetunes open up new possibilities for developers seeking customizable AI without cloud dependency. The abliterated and uncensored variants remove safety alignments, giving users full control over outputs for research, role-play, or creative writing. The 12B parameter scale balances performance and resource requirements, making it accessible for local inference on high-end gaming GPUs or MacBooks. This rapid community adaptation highlights the demand for flexible, open-weight models that can be tailored to specific use cases without policy restrictions.
- Four initial finetunes available: Heretic, Melody1437, Abliterated, and Uncensored
- All models quantized to GGUF for efficient local inference on consumer hardware
- Abliterated and uncensored variants remove safety filters for full output control
Why It Matters
Empowers developers to deploy customized, uncensored Gemma 4 locally for specialized applications.