ByteDance open-sources its 3B-parameter multimodal model, Lance, trained on 128 A100 GPUs

ME News reports that on May 19 (UTC+8), according to monitoring by Beating, ByteDance Research has officially open-sourced Lance, its native unified multimodal large model. This lightweight model features only 3B activated parameters and supports simultaneous understanding, generation, and editing of images and videos within a single framework. Current mainstream unified models heavily rely on scaling up parameter size or adapting text-to-image architectures, whereas Lance successfully demonstrates a low-compute collaborative approach. The development team trained the model entirely from scratch, limiting the total computational budget for the entire training cycle to just 128 A100 GPUs. To address internal conflicts between different modalities and tasks, Lance implements two architectural isolations: - A dual-stream Mixture-of-Experts (MoE) architecture processes interleaved multimodal sequences, decoupling computation paths for understanding and generation while sharing a common underlying context. - Modal-aware rotary positional encoding directly reduces signal interference between heterogeneous visual tokens from images and videos. Extreme computational compression has not compromised performance. With only 3B activated parameters, Lance leads most benchmark tests in image and video generation and editing compared to existing open-source unified models, proving a low-cost path that balances generation capability with semantic understanding through multi-task collaboration. (Source: BlockBeats)