Ai2 Adds On-Demand Embedding Exports to OlmoEarth Studio
Earth-observation teams can generate geospatial vectors for a chosen place and time, then take the resulting raster into their own analysis tools.
Ai2 announced on August 12 that OlmoEarth Studio can now compute and export embedding vectors for a user-selected area, time span and satellite source. The output is a Cloud-Optimized GeoTIFF that teams can move into geospatial software or their own Python workflows instead of keeping the analysis inside the Studio interface.
The change matters to Earth-observation developers because an embedding compresses information about each location into a numerical vector. Similar surfaces should sit closer together in that representation, allowing downstream analysis without training a full task-specific model first. Ai2 presents the export as an entry point for similarity search, lightweight classification, change detection and unsupervised exploration.
Users control place, time, model size and imagery
The Studio workflow lets a user draw or upload an area of interest and choose between one and 12 monthly periods. It offers three OlmoEarth encoder variants: Nano with 128-dimensional output and 1.4 million parameters, Tiny with 192 dimensions and 6.2 million parameters, and Base with 768 dimensions and 89 million parameters.
Spatial resolution can be set to 10, 20, 40 or 80 metres per pixel. The input options are Sentinel-2 L2A optical imagery, Sentinel-1 RTC radar imagery or both. Those controls mean an export can target seasonal change, a specific sensor mix and an output size suited to the user's compute and storage budget.
Ai2 says each embedding dimension becomes one band in the exported raster. The vectors are stored as signed 8-bit integers from -127 to +127, while -128 is reserved for missing data. The project provides a dequantisation route for users who need floating-point vectors in later analysis.
The export opens several downstream paths
For similarity search, a user can select a reference pixel or region and compare its vector with every other location. For classification, a small set of labelled pixels can train a linear model over the exported representation. Monthly exports can be compared to highlight surface changes, while dimensionality reduction can reveal clusters without a predefined label set.
These are flexible building blocks rather than finished monitoring products. A similarity map still needs domain interpretation. A linear classifier is only as reliable as its labels and validation. A change signal can reflect seasonal variation, clouds, sensor differences or data gaps as well as the event a user hopes to detect.
Ai2's demonstrations cover urban surfaces, agriculture, mangroves and a wildfire burn scar. They show how the interface can be used, but they do not establish universal accuracy across regions, seasons or tasks. The source advises users to assess embedding quality for their own application and warns that persistent cloud cover, atmospheric artefacts and missing observations can affect the vectors.
Availability is real but not fully described
Ai2 says custom-computed embeddings are available for OlmoEarth Studio users and asks interested users to make contact. It also says the open-source models can be run independently. The announcement does not specify whether every Studio account receives immediate self-service access, what quotas or prices apply, how long exports take or which regions have been tested.
That makes this a confirmed capability update with an access caveat, not a claim of unrestricted general availability. It is also distinct from Ai2's earlier description of the distributed infrastructure used to run OlmoEarth at continental scale: the new reader value is the portable embedding output and the user-facing configuration surface. The next verifiable checkpoint will be public access terms, operational limits or independent results from teams using exported rasters.
Status
Confirmed. Internal confidence is medium because the feature, formats and demonstrations are described by Ai2 on its verified Hugging Face organisation page without independent operational testing in this run.
Sources
Update note: Last reviewed 2026-08-17. We will revise this post if Ai2 publishes access terms, quotas, pricing or independent evaluations of exported embeddings.
Sources
- Ai2 — Introducing OlmoEarth embeddings — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.