<p>Foundation models, or large atomistic models (LAMs), aim to universally represent the ground-state potential energy surface (PES) of atomistic systems as defined by density functional theory (DFT). Scaling laws suggest that their generalizability improves with increased model size, training data, and computational budgets. We present DPA3, a multi-layer graph neural network founded on line graph series (LiGS), designed for the era of LAMs. We demonstrate that the generalization error of the DPA3 model adheres to the scaling law. The scalability in the number of model parameters is attained by stacking additional layers within DPA3. Additionally, a dataset encoding mechanism decouples training data scaling from model size within a multi-task framework. When trained as problem-oriented potential energy models, the DPA3 model achieves competitive or improved accuracy with substantially fewer parameters than state-of-the-art baselines across molecules, bulk materials, catalysts, two-dimensional materials, and battery materials. When trained as a LAM on the OpenLAM-v1 dataset, the DPA-3.1-3M model achieves strong zero-shot generalization across 12 downstream tasks spanning a diverse array of research domains, demonstrating its potential as an effective out-of-the-box potential model that may require less fine-tuning data for downstream scientific applications.</p>