ByteDance’s Volcano Engine has upgraded Doubao-Seed-2.0-lite, the company’s first full-modal understanding model — meaning it can process video, image, audio, and text all in one. A new Doubao-Seed-2.0-mini with the same capabilities also launched alongside it.
On the vision side, the lite version surpasses February’s Doubao-Seed-2.0-pro in high-level academic benchmarks like physical reasoning (HiPhO) and medical QA (MedXpertQA). It also achieves state-of-the-art results in fine-grained perception (BabyVision, WorldVQA) and embodied understanding (ERQA, which tests a model’s ability to grasp actions and spatial relationships in physical environments). For audio, the model supports speech transcription in 19 languages and translation between 16 languages, beating Gemini 3.1 Pro on speech recognition, translation, and other audio benchmarks. By fusing audio and video, it can jointly analyze what it sees and hears to check consistency. The update arrives as competitors like Mistral push their own powerful models.
Agent and coding capabilities have also been upgraded. The model now supports frameworks like OpenClaw and Hermes Agent, improving multi-step task decomposition and long-horizon stability. Its GUI abilities unify interface recognition and action execution, supporting clicks, typing, scrolling, and drag-and-drop — enabling Browser Use and Computer Use operations for cross-app business workflows.