Gemma 4 abliterated fine-tune strips safety filters, adds image-text-to-text pipeline
DavidAU released a multi-modal Gemma 4 fine-tune this week that removes safety filters and adds image-text-to-text capability, targeting practitioners who need unrestricted inference.
A new Gemma 4 fine-tune promises unrestricted reasoning across text and image inputs. DavidAU's gemma-4-E2B-it-The-DECKARD-Expresso-Universe-HERETIC-UNCENSORED-Thinking is a multi-modal fine-tune of Google's Gemma 4 base model that removes safety guardrails and adds image-text-to-text pipeline support.
The model card tags it as "abliterated" and "uncensored," signaling that content filters present in the base Gemma 4 release have been stripped. The DECKARD Expresso Universe naming suggests a chain of prior fine-tuning steps, though the card does not detail the intermediate checkpoints or training datasets. The model supports any-to-any pipelines, meaning it can ingest combinations of text and image inputs and generate text outputs. No parameter count, context length, or quantization details appear in the available metadata.
The release went live on HuggingFace this week in safetensors format with transformers compatibility, allowing standard loading paths to work without custom code. Practitioners running local inference stacks—ComfyUI, Ollama, LM Studio—can pull the weights directly from the repository. DavidAU has not published a separate model card, paper, or benchmark suite; the HuggingFace tags and naming convention are the only public documentation available.