Overview
- The platform is available now with models, a research paper, documentation, and example deployments, with code and pre‑fine‑tuned models on GitHub and limited gating for some fine‑tuning features.
- OlmoEarth comprises vision‑transformer models trained on roughly 10 terabytes of multimodal, time‑series data that blend optical imagery, radar readings, and contextual maps.
- Companion tools include Studio for no‑code task fine‑tuning, Viewer for interactive map publishing, Run for training management, and APIs for integration, with options to run models offline.
- Early users include IFPRI for crop mapping in Kenya, Amazon Conservation for rapid deforestation detection, Global Mangrove Watch for faster mangrove updates, and a wildfire risk effort with NASA’s Jet Propulsion Laboratory.
- Ai2 reports state‑of‑the‑art results versus established models and offers model sizes from Nano to Large to balance performance with lower inference compute needs.