How Phi-4 Cracked Small Multimodality

Dive into a 36-minute technical video that explores how Microsoft's Phi-4 model achieved multimodal capabilities despite its relatively small size. Learn about the innovative architecture using Mixture of LoRAs that enables Phi-4 to process text, images, and audio efficiently. The presentation walks through the four-stage training process, including vision training, audio training, and reasoning training, with detailed explanations of the model architecture and training data. Discover practical applications for Phi-4, see demonstrations of its capabilities through testing, and understand the technical details behind this breakthrough in small multimodal models. The video includes comprehensive information about the training data, particularly focusing on the visual training data that helped Phi-4 develop its image understanding abilities.