“The Tapestry model will be trained using a federated learning approach where international contributors provide compute and data but only share parameter vectors, not the underlying data itself.”