Aria is an open multimodal native MoE model featuring 3.9B activated parameters per token and a 64K context window. It achieves state-of-the-art performance across language and multimodal tasks, with superior capabilities in video and document understanding. The codebase provides inference and fine-tuning capabilities through Hugging Face Transformers integration, supporting deployment on a single A100 GPU with bfloat16 precision.