Guilherme34/Firefly-v4
Guilherme34/Firefly-v4 is a 5.1 billion parameter uncensored, multimodal roleplay and creative writing model fine-tuned from gemma-4-E2B-it-heretic-ara. It natively supports vision and audio inputs, allowing it to process images and sound within conversations. A key feature is its switchable reasoning, enabling users to toggle chain-of-thought processing for either deeper coherence or faster, more spontaneous output. This model is designed for immersive, unfiltered roleplay and creative tasks without built-in refusals or alignment guardrails.
Loading preview...
Firefly-v4: Uncensored Multimodal Roleplay and Creative Writing
Firefly-v4 is a 5.1 billion parameter model developed by Guilherme34, fine-tuned from p-e-w/gemma-4-E2B-it-heretic-ara using Unsloth. It is specifically designed for uncensored, multimodal roleplay and creative writing, offering full creative freedom without refusals or character breaks.
Key Capabilities & Features
- Multimodal: Natively supports vision and audio inputs, allowing it to understand images and process audio as part of a conversation, not as an external pipeline.
- Uncensored: Provides an unfiltered experience for roleplay and storytelling, free from alignment guardrails or moralizing.
- Switchable Reasoning: Users can toggle chain-of-thought processing on or off using the
<|think|>tag in the system prompt. This allows for either enhanced coherence in complex scenarios or faster, more direct responses. - Creative Focus: Excels at immersive roleplay and general creative tasks, as well as instruction following.
Training Details
Firefly-v4 was fine-tuned from a Gemma 4 base model, inheriting its native text, vision, and audio modalities. The training focused on uncensored roleplay, creative writing, and hybrid reasoning capabilities.
License
This model operates under the Firefly-v4 Attribution License. It is free for personal and non-commercial use. Commercial use requires prominent attribution to "Based on Guilherme34/Firefly-v4", including the model name and author, with a link to the original HuggingFace repository if possible. Derivatives must also carry forward these attribution terms.