Distributed applications often eschew strong consistency and replicate data asynchronously to improve availability and fault tolerance. However, programming under eventual consistency is significantly more complex and often leads to onerous programming model where inconsistencies must be handled explicitly. We introduce vml, a programming model that extends ML datatypes with mergeability a la version control systems with the ability to define and compose distributed ML computations around such data. Our OCaml implementation instantiates mergeable types on Irmin, a distributed content-addressible store to enable composable and highly-available distributed applications.