Unified Multimodal AI Model Advances Scientific Discovery.

Scientific research is increasingly moving towards multi-domain reasoning, and AI models need to adapt to this shift. Current AI systems for science often specialize in single domains or rely heavily on text-based interfaces, limiting their ability to process diverse scientific data and generate modality-native outputs. To address this, a new model called Monkey King Bang (MKB) has been developed.MKB is a unified scientific multimodal foundation model designed for both understanding and generation across a broad spectrum of scientific disciplines. It incorporates a shared Transformer backbone, complemented by specialized encoders, adapters, and decoders tailored for different modalities. The model covers six key scientific branches: DNA, RNA, proteins, small molecules, earth science, and medical images, supporting native outputs like biological sequences, molecular strings, meteorological fields, and segmentation masks.The training process for MKB involves a two-stage curriculum, first aligning modality-specific components with a frozen backbone, then consolidating them with a language backbone using mixed scientific and general corpora. Experimental results show MKB's competitive performance in scientific understanding across various benchmarks and its ability to produce high-fidelity native outputs for tasks such as weather forecasting and medical image segmentation, while largely retaining the general capabilities of its base Qwen3-VL model. This demonstrates the potential of shared-backbone models with tailored components for cross-domain scientific exploration.