Allen Institute for AI released OlmoEarth v1.3 together with precomputed annual embeddings for the whole planet and a browser-based map-making tool. Each 10-by-10-meter patch of ground gets a numerical fingerprint that compresses a year of satellite observations, with coarser resolutions also available. The release removes the previous requirement to fine-tune a separate model and run GPU-heavy processing over raw imagery. For business users, this matters because testing a land-monitoring idea now takes seconds instead of hours.
Embeddings Explorer and cheaper v1.3 training
OlmoEarth Embeddings Explorer lets users turn embeddings into working maps without sign-in, from tracking change over time to finding similar places. Previously teams had to fine-tune in OlmoEarth Studio, train a model for each task, download imagery and secure large GPU capacity. Ai2 said testing an idea previously meant an expensive run on the order of hours, while a first pass in the explorer takes seconds. Analysis of a region from a handful of labels requires no GPU at all.
Version 1.3 reads a year of radar and optical data from Sentinel-1, Sentinel-2 and Landsat. Sentinel-1 measures radar reflections, while Sentinel-2 and Landsat capture visible and infrared wavelengths. The model distills patterns learned from millions of images into fingerprints that other software can use to separate crops, forests, mines and other land cover. Ai2 described the embeddings as AI-ready inputs for downstream mapping without additional model training.
Training efficiency is a central change in this release. Version 1.3 took 53% of the computing time of v1.2 and 18% of the computing time of v1, according to Ai2. Carbon emissions were 50% of v1.2 and 22% of v1. Ai2 called v1.3 its cheapest model to train, while noting that v1.2 remains valuable for research and partner labs as a strong fine-tuning model.
What planetary embeddings change for companies
For companies that monitor land, supply chains or environmental exposure, precomputed planetary coverage lowers the entry cost of a pilot. An agricultural analyst can map oil palm cultivation and flag areas where expansion may have replaced forest, then decide where closer examination is warranted. Small firms gain because they can test hypotheses in a browser without imagery storage or compute contracts. Large organizations gain faster triage before committing budgets to full fine-tuning.
The Peru example shows how year-to-year comparison works without training. Illegal gold mining around La Pampa in Madre de Dios peaked in 2017 and 2018, clearing more than 1,200 hectares in the buffer zone of Tambopata National Reserve, as reported by the MAAP monitoring program. After Operation Mercurio in February 2019, embedding comparisons for 2016 to 2020 showed 34 square kilometers of significant change in the southern mining belt. The same comparison for 2020 to 2024 falls to 9 square kilometers, making the shutdown visible.
The marker to watch is whether teams publish operational maps built only from explorer outputs and handfuls of labels. If insurers, traders, forestry operators or auditors start using embedding comparisons for routine screening, then planetary fingerprints have moved from research input to production data layer. Continued releases of annual planet-wide embeddings at 10-meter detail would confirm that direction.
