ReleasesNSFW
DavidAU drops two abliterated Gemma-4 E2B multimodal fine-tunes on HuggingFace
Two uncensored Gemma-4 E2B fine-tunes from DavidAU hit HuggingFace on May 19, both built for image-text-to-text workflows with safety layers removed.
DavidAU released two uncensored Gemma-4 E2B fine-tunes on HuggingFace on May 19, both carrying the HERETIC designation and built on the any-to-any pipeline. The models — gemma-4-E2B-it-V2-The-DECKARD-Expresso-ONE-Universe-HERETIC-UNCENSORED-Thinking and its V2b sibling — are tagged as abliterated fine-tunes with no safety filtering.
Both checkpoints support image-text-to-text generation and ship as safetensors weights. The model cards list transformers compatibility and the gemma4 base architecture.
What stands out
- 01Abliterated base — Both models carry the abliterated tag, signaling that safety layers from the original Gemma-4 E2B instruction-tuned checkpoint were stripped during the fine-tune process.
- 02Multimodal pipeline — The any-to-any pipeline tag indicates support for image-text-to-text workflows, not just text completion. Practitioners can feed images and text prompts together.
- 03DECKARD-Expresso-ONE-Universe naming — The verbose model names suggest a multi-stage fine-tuning recipe. DavidAU has released other DECKARD and Expresso variants in the past, typically combining multiple LoRA merges or sequential training runs.
- 04V2 vs V2b split — The two releases differ only by a single character in the version string. The model cards do not specify what changed between V2 and V2b; practitioners will need to test both to see if one outperforms the other on specific prompts.