128 A100 trained from scratch! Byte Open Source 3B Omnipotent Multimodal Model Lance According to the dynamic observation of Beating, ByteDance Research officially opened the original unified multi-modal large model Lance. This is a lightweight model with only 3B activation parameters, which supports the understanding, generation, and editing of both images and videos within a single framework. At present, the mainstream unified model is highly dependent on expanding the parameter scale or following the Wensheng map architecture, while Lance runs through the collaborative route of extremely low computing power. The R&D team trained the model completely from scratch and reduced the total computational budget for the entire training cycle to 128 A100 GPUs. To address internal conflicts between different modalities and tasks, Lance implemented two hard isolations in the architecture: - using a dual stream hybrid expert (MoE) architecture to handle interwoven multimodal sequences, decoupling the understanding and generation of computational paths while sharing underlying context. -Introducing modal aware rotational position encoding directly reduces signal interference between heterogeneous visual tokens in images and videos. Extreme computing power compression has not lowered the performance limit. When there are only 3B activation parameters, Lance's image and video generation and editing performance leads the existing open source unified model in most benchmark tests, and through multi task collaboration, it runs a low-cost route of giving consideration to both generation and semantic understanding of small parameters. [Original link]