OlmoEarth by Ai2

OlmoEarth by Ai2 is an analytics tool. OlmoEarth is Ai2's geospatial AI platform and open model family for satellite data. It covers fine-tuning, embeddings and production inference, aimed at conservation, climate and research teams that lack deep AI expertise. Access is by request.

OlmoEarth is a geospatial AI offering from Ai2, the non-profit research institute founded by Paul Allen. It combines a family of foundation models trained on millions of global Earth observations with a hosted platform that handles the work between raw satellite data and a usable result. The stated aim is to put advanced remote-sensing AI in the hands of organizations and communities that lack in-house machine learning expertise. In practice, the platform covers the full path from data to deployment. Users start with Earth data, run research and development, fine-tune a model for their own question, export custom embeddings from OlmoEarth Studio for downstream analysis, and then push the result into production. The engineering blog describes continent-scale satellite inference as a design target, and the homepage says work that once took months of processing can now finish in hours, which makes recurring monitoring feasible on modest budgets. A documented API is available for programmatic access, and the v1.1 model family is described as cutting compute costs by up to 3x. The homepage leans on partner stories to show where it is used. Global Mangrove Watch produced more accurate maps with about a tenth of the data points previously needed. NASA JPL and Google.org are using it to build a next-generation fire risk model. The IUCN points to OlmoEarth foundation models running inside Skylight, an ocean intelligence platform used for high seas protection. Conservation groups, environmental agencies, and applied researchers are the clearest audience. Among alternatives, OlmoEarth sits between closed commercial geospatial analytics services and do-it-yourself research code. Researchers can download the open models and the pretraining dataset from Hugging Face and the model code from GitHub, while teams that want a managed route request an account on the platform. That open-plus-hosted split is the main differentiator against vendors that only sell outputs. Access is currently by request, so availability is gated rather than self-serve.