Alduin 4B: Uncensored Vision LLM Based on Gemma 3 Released on Hugging Face
A 4B-parameter model that sees images and refuses no prompts—open and uncensored.
Muhammadreza, known on Reddit as Haghiri75, has released Alduin 4B, a 4-billion-parameter vision-language model that is fully uncensored and open-source. Built on Google's Gemma 3 architecture, the model can process both text and images, making it a multimodal tool for local, unrestricted use. The author's explicit goal is to give users the right to use the model in any way they choose, without safety filters or content restrictions. The model is hosted on Hugging Face at https://huggingface.co/Muhammadreza/alduin-4b-it-base, and while no quantized version is currently available, the author welcomes community contributions for quantization.
The release targets developers and power users who want complete control over AI behavior—particularly for tasks where censorship or ethical guardrails are seen as limitations. By leveraging the lightweight Gemma 3 base, Alduin 4B can run locally on modest hardware, yet still handle vision tasks alongside text generation. The model's uncensored nature means it will comply with any prompt, including those that other models might refuse, making it a niche tool for experimental or edge-case applications. This release underscores the growing tension between safety-constrained models and the open-source community's demand for unrestricted access.
- Alduin 4B is a 4-billion-parameter vision-language model based on Google's Gemma 3, capable of understanding both text and images.
- Released as an uncensored model—no safety filters—allowing full freedom in how it is used.
- Available on Hugging Face with no quantized version yet; the author encourages community quantization contributions.
Why It Matters
Offers a locally-run, completely unrestricted vision LLM, appealing to developers who reject AI safety constraints.