Gemma 4 abliterated fine-tune adds multimodal reasoning to uncensored weights

DavidAU released an uncensored Gemma 4 fine-tune this week that strips safety filters from Google's base model while adding multimodal image-text-to-text capability. The checkpoint, named gemma-4-E2B-it-The-DECKARD-Expresso-ONE-Universe-HERETIC-UNCENSORED-Thinking, uses abliteration—a technique that removes safety refusal behavior from model weights—and ships in safetensors format for local inference on ComfyUI, Ollama, and similar toolchains. The model card tags it "heretic" and "uncensored," signaling unrestricted prompting for practitioners.

The DECKARD Expresso naming suggests a custom thinking layer or reasoning scaffold on top of Gemma 4, though the model card does not detail the training recipe or parameter count. Abliteration has become standard practice in the open-weight community, targeting the refusal neurons that cause models to decline certain prompts. The technique originated with Llama fine-tunes and has since spread to Qwen, Mistral, and now Google's Gemma family. The any-to-any pipeline tag indicates the model handles multiple input and output modalities; image-text-to-text capability means users can feed an image alongside a text prompt and receive a text response, useful for captioning, visual question answering, and multimodal reasoning.

Google released Gemma 4 as an open-weight multimodal model earlier this year with built-in safety tuning that blocks adult content, violence, and other restricted categories. Third-party fine-tuners routinely strip those guardrails for use cases where content filtering is undesirable—creative writing, uncensored chatbots, and synthetic data generation. The checkpoint went live on HuggingFace on May 18, 2026.