Moondream2 is a 1.86 billion parameter model initialized with weights from SigLIP and Phi-1.5. This compact architecture allows for efficient processing while maintaining robust capabilities.
Designed to run on devices with low-resource settings, Moondream2 optimizes memory usage and processing power. This makes it ideal for deployment on smartphones, IoT devices, and other edge computing scenarios.
Evaluated on various tasks including table, form, and complex document understanding, Moondream2 shows promising results for a small model. It can extract key information from diverse document types with impressive accuracy.