Lance: Unified Multimodal Modeling by Multi-Task Synergy

In three linesLance is a lightweight unified multimodal model supporting understanding, generation, and editing of images and videos. Built on dual-stream mixture-of-experts architecture with modality-aware rotary positional encoding, it combines collaborative multi-task training and adaptive data scheduling to outperform existing open-source unified models in visual generation.

Summary generated by Claude — human-verified