Five Gemma-4 12B abliterated checkpoints land on HuggingFace in GGUF and MLX
Five Gemma-4 12B uncensored fine-tunes appeared on HuggingFace on June 21–22, spanning chain-of-thought, coder, and instruct variants in GGUF and MLX quantizations.
Five uncensored Gemma-4 12B checkpoints landed on HuggingFace on June 21–22, 2026, each targeting different use cases and hardware profiles. The releases include chain-of-thought reasoning builds, coder-focused variants, and general instruction-tuned models, all stripped of safety guardrails.
Rangle2's gemma-4-12B-uncensored-opus4.7-cot pairs abliteration with chain-of-thought prompting and multimodal image-text-to-text capability. Nightmedia shipped two MLX-quantized builds: gemma-4-12B-coder-fable5-composer2.5-v1-uncensored-heretic-mxfp8-mlx and gemma-4-12B-it-uncensored-heretic-mxfp8-mlx, both in mxfp8 precision for Apple Silicon. The coder variant merges Fable5 and Composer 2.5 training, while the instruct-tuned build targets general chat.
Quantization formats
Mradermacher contributed two GGUF conversions of the nightmedia coder checkpoint: gemma-4-12B-coder-fable5-composer2.5-v1-uncensored-heretic-GGUF and an i1-quantized variant. Both target llama.cpp and similar CPU/GPU inference engines, with emphasis on coding, reasoning, and thinking tasks.
All five checkpoints carry the "uncensored" and "abliterated" tags, signaling removal of refusal behavior. The simultaneous drop across multiple quantization formats—MLX for Apple Silicon, GGUF for CPU/GPU inference, and native safetensors—reflects coordinated community interest in running Gemma-4 12B without content filters on consumer hardware.
