The OlmoEarth Platform: Geospatial inference at planetary scale

Details

Allen Institute for AI (Ai2) has released OlmoEarth Platform, a managed fine-tuning and inference layer built around its OlmoEarth family of Earth observation foundation models. The models were pretrained on roughly 10 terabytes of multimodal satellite data. Governments, NGOs, and environmental organizations are already deploying them for deforestation monitoring, food security assessment, and wildfire risk mapping. The platform's headline number is cost-efficient continent-scale inference: a full continent can be processed in roughly a day, handling dozens of terabytes of imagery at fractions of a penny per square kilometer. Technically, the system handles the hard parts of satellite inference — aligning imagery across providers with incompatible projections and resolutions, processing multi-spectral bands, and stitching outputs into geographically consistent maps across distributed computing failures. The motivation is explicitly practical. Ai2 already open-sources the underlying model weights, which works fine for organizations with ML engineering capacity. But most environmental organizations that most need geospatial AI cannot manage the full pipeline: labeling, fine-tuning, and large-scale inference. OlmoEarth Platform is the managed layer bridging that gap, informed by Ai2's decade running operational platforms like Skylight and EarthRanger. For practitioners, this is a notable pattern: domain-specific foundation models paired with managed deployment infrastructure, specifically designed for mission-driven organizations with limited engineering bandwidth.