This paper highlights that next-generation multimodal-based models capable of all types of two-way and multi-turn interactions will become a core element of artificial general intelligence systems. To overcome the limitations of existing models, this paper introduces NExT-OMNI, an open-source omnimodal-based model that achieves unified modeling through the discrete flow paradigm. NExT-OMNI leverages metric-guided probabilistic paths and optimal motion velocities to support all types of understanding and generation. Its concise, unified representation, rather than a task-separated design, makes it applicable to a wide range of scenarios. Trained on large-scale text, image, video, and audio data, NExT-OMNI demonstrates competitive performance on multimodal generation and understanding benchmarks and outperforms previous unified models in multi-turn multimodal interaction and cross-modal retrieval. To facilitate research, we open-source the training details, data protocol, code, and model checkpoints.
