The disclosure generally describes computer-implemented methods, computer program products, and systems for providing metadata-driven dynamic load balancing in multi-tenant systems. A computer-implemented method includes: identifying a request related to a model-based application executing in a multi-tenant system associated with a plurality of application servers and identifying at least one object in the model-based application associated with the request. At least one application server is identified as associated with a locally-cached version of a runtime version of the identified object, and a determination of a particular one of the identified application servers to send the identified request for processing is based on a combination of the availability of a locally-cached version of the runtime version at the particular application server and the server's processing load. The request is then sent to the determined application server for processing.