Allen AI released OlmoEarth embeddings capability in OlmoEarth Studio, allowing users to compute and export compact numerical representations of Earth-observation data from open-source OlmoEarth foundation models. The Studio platform now supports three encoder variants—Nano (128-dim, 1.4M params), Tiny (192-dim, 6.2M params), and Base (768-dim, 89M params)—each tunable for spatial resolution from 10 to 80 meters per pixel. Embeddings are exported as Cloud-Optimized GeoTIFFs (COGs), lightweight and easy to share.
OlmoEarth embeddings enable fast, cost-effective downstream tasks without full model inference: similarity search to find visually similar landscape patches, few-shot segmentation from as few as 60 labeled pixels (20 per class), change detection across monthly time spans, and unsupervised exploration. Users can query a pixel, compute cosine similarity, or train a simple linear classifier on frozen embeddings with strong results—achieving weighted F1=0.84 on mangrove classification from just 60 labeled pixels. Everything is computed on-demand, reflecting exactly the conditions users care about.
For geospatial and climate teams, OlmoEarth embeddings unlock analysis at planetary scale without massive compute bills. Allen AI has spent over a decade operating platforms like Skylight and EarthRanger, so this production-ready tooling matters. The embedding-first approach is crucial for NGOs, governments, and enterprises monitoring deforestation, food security, wildfire risk, and environmental change. The shift from open models to a platform with embeddings, fine-tuning, and fleet inference suggests earth observation is maturing as a category where both model capability and operational accessibility are required.