Ryex/Floppa_Gemma4-E4B-it

VISIONConcurrency Cost:1Model Size:7.9BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Ryex/Floppa_Gemma4-E4B-it is a 7.9 billion parameter multimodal model developed by Ryex, based on the Gemma 4 E4B architecture. Fine-tuned for general usage, its primary differentiator is the complete removal of refusal patterns, enabling unrestricted image description of explicit or violent content and uncensored translation of raw media. It maintains multimodal vision capabilities with a 32768 token context length, making it suitable for specialized translation and content analysis tasks where content filtering is undesirable.

Loading preview...

Floppa_Gemma4-E4B-it: Uncensored Multimodal AI

Developed by Ryex, Floppa_Gemma4-E4B-it is a 7.9 billion parameter multimodal model built upon the Gemma 4 E4B architecture, specifically fine-tuned from llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic. Its core objective is extreme behavior modification, focusing on the complete removal of refusal guardrails while preserving its multimodal vision capabilities. This model was trained using 16-bit QLoRA via Unsloth and Huggingface's TRL library, leveraging a balanced dataset of approximately 10.5k rows.

Key Capabilities

  • "Spicy" Vision: Describes NSFW, violent, or explicit imagery (anime, photography) without safety refusals.
  • Uncensored Translation: Translates dialogue, including slurs, profanity, and aggressive language, exactly as written, ideal for localizing raw media.
  • Multimodal Context: Utilizes Gemma 4's architecture to translate text within images and provide context-aware scene descriptions for complex translation tasks.

Training & Optimization

The model's fine-tuning dataset, dubbed the "Floppa Mix," includes 20% toxic/uncensored text, 20% high-quality multilingual pairs (Unbabel/TowerBlocks-v0.2), 40% general reasoning (mlabonne/FineTome-100k), and 20% vision anchors (merve/vqav2-small + custom anime dataset) to prevent visual catastrophic forgetting. It is optimized for deployment with vLLM and Text-Generation-Inference.

Good For

  • Research into uncensored language models.
  • Specialized translation tasks requiring raw, unfiltered output.
  • Creative writing workflows where content filtering is undesirable.
  • Analysis of explicit or sensitive visual content without AI-imposed restrictions.

Finetunes
2 models