Lance: Unified Multimodal Modeling by Multi-Task Synergy

Lance is a unified multimodal model trained from scratch using a dual-stream mixture-of-experts architecture on interleaved sequences, supporting image and video understanding, generation, and editing jointly. It prioritizes multi-task synergy over scale, offering a practical alternative to modality-siloed designs, though no benchmark numbers are provided in the excerpt.