Polymathic AI develops and releases open foundation models and large-scale datasets for machine learning on scientific data, focusing on numerical simulations and multimodal observations from physics, astrophysics, and biology. Releases include Walrus, a 1.3 billion-parameter transformer-based model trained across diverse physical systems for continuum dynamics; AION-1, a billion-parameter multimodal foundation model unifying heterogeneous astronomical data from images, spectra, and time series; MIMIC, a generative multimodal model jointly modeling DNA, RNA, and proteins; and Overtone, a compute-flexible transformer for PDE-based physics emulations. Datasets provided are The Well, containing 15TB of spatiotemporal physics simulations, and The Multimodal Universe, with 100TB of standardized astronomy data. These outputs support the scientific community in modeling complex phenomena. The initiative is led by principal investigator Shirley Ho, group leader at the Flatiron Institute’s Center for Computational Astrophysics. It maintains offices in Manhattan, New York, NY.