In May 2026, a team from the University of Adelaide in partnership with the SmartSat Cooperative Research Centre successfully deployed Prithvi — the open-source geospatial foundation model jointly developed by NASA and IBM — onto two on-orbit platforms: the South Australian government's Kanyini satellite and the International Space Station's IMAGIN-e payload, making it the first geospatial AI foundation model running in orbit.
Prithvi is trained on 13 years of Landsat and Sentinel-2 satellite remote-sensing data, with cross-task generalization capability that can adapt to flood monitoring, cloud identification, crop yield estimation, and other downstream tasks. The core breakthrough: traditional satellites are bandwidth-limited, making it nearly impossible to upload new models, while Prithvi only needs a small decoder extension pack added to complete task adaptation, dramatically lowering the barrier to on-orbit updates.
The open-source strategy is equally key. Research lead Dr. Andrew Du stated bluntly: "If Prithvi weren't open-source, I'd have to train my own foundation model from scratch." NASA and IBM's open route lets global researchers quickly reuse and innovate, validating the value of the foundation model concept.
NASA also plans to release more foundation models in planetary science, astrophysics, and other domains, continuing to expand the boundaries of AI for Science.