The foundation model paradigm has revolutionized NLP and CV, yet its application to 3D Partial Differential Equations (PDEs) remains constrained by computational bottlenecks, inconsistent generalization, and a lack of multi-task versatility. Current PDE models often function as sophisticated weight initializers rather than true foundation models, frequently requiring expensive full-parameter fine-tuning to adapt to new physics. Furthermore, the high cost of generating and storing 3D data has largely restricted research to 1D and 2D domains. In this work, we present Tadpole (Three-dimensional autoencoders for PDEs with online learning), a novel foundation model designed for high-dimensional physical systems. Tadpole overcomes the data bottleneck through an online learning framework featuring GPU-based solvers and a specialized buffer strategy, enabling training on hundreds of terabytes of data by bypassing traditional I/O and storage limits. By pre-training as an autoencoder on single-channel crops, Tadpole learns highly transferable latent representations across diverse PDE systems and resolutions. To ensure efficient adaptation, we introduce CoBaLT, a Parameter-Efficient Fine-Tuning (PEFT) method that leverages the Tadpole backbone to significantly reduce dynamics learning costs without sacrificing accuracy. Our results demonstrate that Tadpole is a truly multi-functional foundation model, excelling in autoencoding, dynamics learning, and generative modeling for 3D PDEs with the same pretrained model. This framework provides a scalable and accessible pathway for applying large-scale foundation models to complex, real-world scientific problems.