Gemma 4 Ortenzya 31B uncensored weights land on HuggingFace in MLX-8Bit format
McG-221 released an 8-bit MLX-quantized version of the 31-billion-parameter Gemma 4 Ortenzya model, tagged uncensored and optimized for Apple Silicon inference.
Gemma 4 Ortenzya The Creative Wordsmith 31B is a multimodal image-text-to-text model from McG-221, now available on HuggingFace in 8-bit MLX-quantized format. The checkpoint is tagged "uncensored" and "heretic," signaling removal of safety filters, with safetensors weights optimized for Apple Silicon via the MLX framework. The 31-billion-parameter base supports both text generation and image-conditioned prompting. MLX-8Bit quantization typically cuts memory requirements by roughly half compared to FP16, bringing a 31B-parameter model within reach of unified-memory Macs with 64GB or more.
The model card lists Unsloth and Text Generation Inference compatibility, suggesting the weights can slot into standard transformer pipelines for chat or creative writing tasks. No benchmark numbers or training methodology appear in the listing. As of mid-May 2026, the repository shows 60 downloads and 1 like.