Aria

Aria is an open-source multimodal native mixture-of-experts (MoE) AI model developed by Rhymes AI, designed to handle text, images, and video inputs natively. It delivers state-of-the-art performance across a wide range of language and vision benchmarks while remaining efficient through its sparse MoE architecture. The model is publicly available on GitHub and Hugging Face, making it accessible for researchers and developers who want to build or fine-tune multimodal AI applications.

• Multimodal Native Architecture: Aria processes text, images, and video in a unified model without relying on separate encoders bolted together, enabling richer cross-modal understanding.
• Mixture-of-Experts (MoE) Design: Uses a sparse MoE approach so only a subset of parameters are activated per token, delivering high capability with lower inference cost.
• Open-Source Access: The full model weights and code are released on GitHub and Hugging Face under an open license, allowing anyone to download, run, and fine-tune the model.
• State-of-the-Art Benchmarks: Achieves competitive or leading results on standard language, vision-language, and video understanding benchmarks.
• Fine-Tuning Support: Includes scripts and documentation for supervised fine-tuning (SFT) on custom datasets, enabling domain-specific adaptation.
• Inference Recipes: Provides ready-to-use inference code and examples for running the model locally or on cloud GPU infrastructure.
• Community-Driven Development: Hosted on GitHub, the project welcomes contributions, issue reports, and pull requests from the broader AI research community.