The Allen Institute for AI (AI2) released OlmoEarth, a family of open-source multimodal foundation models for Earth observation and geospatial intelligence. The model family comes in four sizes: OlmoEarth-Nano (~1.4M parameters) and Tiny (~6.2M) for fast, cheap inference; Base (~90M) for balanced accuracy/speed; and Large (~300M) for peak performance. Trained on multimodal satellite data (optical, radar, environmental maps), OlmoEarth excels at scene classification, semantic segmentation, object detection, change detection, and regression on single-image and time-series tasks.
OlmoEarth outperforms larger models including Meta's DINOv3, IBM/NASA's Prithvi, IBM's Terramind, CROMA, Panopticon, and AI2's own Satlas and Galileo. It competes with Google DeepMind's AlphaEarth Foundations: on kNN tasks it performs on par or better; fine-tuned, it substantially outperforms. Real-world impact is immediate: Global Mangrove Watch reduced their mapping pipeline from years of data collection/annotation to hours using OlmoEarth, using just 0.1% of their 5.8M labeled samples. The International Food Policy Research Institute is now generating countywide crop maps for Kenya with local field data, enabling faster interventions on food security.
For climate/environment AI architects: OlmoEarth Platform is an integrated end-to-end system (Studio for labeling/fine-tuning, Viewer for exploration, Run for large-scale inference, APIs for workflow integration). The open weights + code + technical paper enable transparency and community audit. Key advantage: smaller models achieve Pareto-optimal performance vs. compute, reducing inference costs and enabling continual re-analysis. Addressable use cases span crop monitoring, mangrove conservation, wildfire risk prediction, and deforestation detection—all requiring frequent refreshes at scale.