Ai2 releases OlmoEarth v1.3 with open 10-meter satellite embeddings and a browser explorer for mapping Earth changes

Ai2 released OlmoEarth v1.3, an open satellite embedding model that turns annual Sentinel-1, Sentinel-2, and Landsat time series into 128-dimensional vectors for every 10-meter pixel, with a lighter 64-dimensional option. In 27 evaluations, the 128-dimensional embeddings ranked highest on 20 tests, while the 64-dimensional version led on 18; Ai2 also released precomputed embeddings, model weights, data, code, a technical report, and the OlmoEarth Embeddings Explorer. At launch, 2025, 2024, and 2023 embeddings are global, while 2017 to 2022 cover select countries, with more coverage planned.

Teams building land-cover or change-detection workflows can use the browser explorer and precomputed embeddings to test ideas in seconds before committing to fine-tuning runs that require GPUs and imagery storage. The trade-off is that the reusable embeddings lower compute and setup costs, but projects needing a specific time window, full historical coverage, or task-specific accuracy may still need OlmoEarth Studio or custom processing; each annual archive is also about 150 TB, so bulk use remains a data-management decision.