ARIA: An Open Multimodal Native Mixture-of-Experts Model

Aria is an open multimodal native mixture-of-experts model introduced by a team of researchers led by Dongxu Li. This model aims to address the limitations of proprietary multimodal AI models by providing an open-source alternative that integrates diverse modalities for comprehensive understanding. Aria boasts impressive performance metrics, with 3.9 billion and 3.5 billion activated parameters for visual and text tokens, respectively. It outperforms existing models such as Pixtral-12B and Llama3.2-11B, and competes effectively with top proprietary models across various multimodal tasks. The model was pre-trained from scratch using a four-stage pipeline that enhances its capabilities in language understanding, multimodal comprehension, long context processing, and instruction following. The researchers have made the model weights and a supporting codebase available for public use, facilitating easy adoption and adaptation in real-world applications.

- Aria is an open-source multimodal AI model designed to integrate diverse information modalities.

- It features 3.9 billion and 3.5 billion activated parameters for visual and text tokens, respectively.

- The model outperforms several existing proprietary models in various multimodal tasks.

- Aria was pre-trained using a four-stage pipeline to enhance its capabilities.

- The model weights and codebase are available for public use, promoting accessibility and adaptability.