Gemma-4 31B abliterated checkpoints optimized for Apple Silicon debut on HuggingFace
Two uncensored Gemma-4 31B fine-tunes packaged in 8-bit MLX format for local inference on M-series Macs appeared on HuggingFace this week.
Two uncensored Gemma-4 31B checkpoints surfaced on HuggingFace this week, both converted to 8-bit MLX format for Apple Silicon inference and tagged with the "heretic" abliteration marker.
G4-MeroMero-31B-uncensored-heretic-mlx-8Bit is the first, a decensored Gemma-4 fine-tune packaged as safetensors weights. The model card lists "abliterated" and "ara" among its tags, indicating safety-layer removal and potential anime-style training bias. A second checkpoint, gemma-4-Ortenzya-The-Creative-Wordsmith-31B-it-uncensored-heretic-mlx-8Bit, carries the same uncensored treatment with an image-text-to-text pipeline tag and an Unsloth conversion marker. Both models showed zero downloads and zero likes at the time they went live on May 17, suggesting fresh releases.
The MLX 8-bit quantization targets MacBook Pro and Mac Studio users running Apple's MLX framework, which ships native GPU acceleration for M-series chips. Gemma-4 31B at full precision requires roughly 62 GB of VRAM; 8-bit quantization cuts that to under 16 GB, making the model runnable on a 64 GB unified-memory M3 Max or M4 Max without swapping to disk. The "heretic" tag is a community convention for abliterated weights — models that have had refusal layers surgically removed via activation engineering or fine-tuning on uncensored datasets.
Both checkpoints are hosted under the cookietimeh namespace. The Ortenzya variant's "Creative Wordsmith" subtitle and Italian "it" tag suggest a creative-writing or roleplay fine-tune, though the model card offers no specifics on the base model or training corpus.
