Gemma 3 12B abliterated weights arrive in fp8 quantization

DreamFast-gemma-3-12b-it-heretic-v2-fp8, released by ibyteohdear on HuggingFace on June 25, is an abliterated variant of Google's Gemma 3 12B Instruct model. The checkpoint removes built-in safety filters and ships in fp8 quantization, cutting memory overhead while preserving the base model's instruction-following capabilities. Abliteration rewrites refusal pathways in the weights rather than prompting around them—a technique that has gained traction among practitioners running local inference stacks who want deterministic unrestricted output without per-query jailbreak gymnastics.

The fp8 format targets consumer GPUs with 24 GB VRAM or less, making the 12-billion-parameter checkpoint accessible to hobbyists and small labs running inference on RTX 4090 or equivalent hardware. The model supports text-generation pipelines and carries tags for image-text-to-text workflows, though the model card does not specify whether visual capabilities were added in the heretic fork. Gemma 3 12B's native architecture handles extended context and multi-turn dialogue; this variant carries those traits into unrestricted territory.

Abliteration has become standard post-training practice in the open-weight community. Where earlier jailbreak methods relied on prompt engineering to bypass safety layers at inference time, abliteration surgically removes the refusal behavior from the weights themselves. The result is a model that behaves as if it was never safety-tuned, without the brittleness of prompt-based workarounds. Google's Gemma family has seen multiple abliterated forks since the original 2B and 7B releases; the 12B instruct variant offers a middle ground between the smaller models' speed and the 27B checkpoint's capability ceiling. The "heretic-v2" naming suggests this is at least the second iteration of ibyteohdear's abliteration pipeline for Gemma 3 12B.