The Allen Institute for AI (Ai2) released OlmoEarth v1.3 on October 22, 2026 — an open satellite-embedding model with precomputed annual AI “fingerprints” for the entire planet at 10-meter resolution, plus a free browser-based Embeddings Explorer that needs no sign-in. The release is published on Hugging FaceIf you’ve ever wanted to analyze the whole Earth from a laptop, this is the moment.
What OlmoEarth v1.3 actually is
OlmoEarth is an embedding model: a 128-million-parameter vision transformer that reads a full year of time series — Sentinel-1 radar, Sentinel-2 and Landsat visible and infrared — and compresses it into a 128-number fingerprint for every 10-by-10-meter patch of Earth. Those fingerprints capture patterns that other software can use to tell crops from forests or mines from cities. Until now, using OlmoEarth meant fine-tuning a model per task in OlmoEarth Studio, which demanded heavy GPU time plus downloading and storing imagery.
The big change: compute once, reuse everywhere
Ai2 computed the embeddings for the entire planet and shared them openly on Source Cooperative — 2023, 2024 and 2025 are global, while 2017 through 2022 cover selected countries and are being backfilled. From there, making a crop map, a land-cover map, or a change analysis becomes a lightweight modeling problem: train a small classifier on top and run it locally, with no imagery downloads and no GPU farm. And the new OlmoEarth Embeddings Explorer runs right in the browser — make a first map or compare places in seconds, track change over time, find similar places — with no sign-in and no advanced technical knowledge required.
Cheaper to train, greener too
Ai2 says v1.3 took 53% of the computing time of v1.2 and 18% of the very first version, with 50% and 22% of the carbon emissions — while remaining a strong fine-tuning model for research labs that want to go deeper. It's another example of
AI for science getting more efficient, not just more powerful.
What you can actually do with it
Deforestation and change tracking, crop mapping, land-cover classification, spotting illegal mining. Ai2's own demo compared the Madre de Dios region of Peru across 2016–2020 and 2020–2024: 34 square kilometers of significant change in a mining belt in the first period, 9 square kilometers in the second — presented as Ai2's demonstration, not an independently verified measurement. Honest caveat: these are embeddings — ready-to-use features, not finished maps — and deeper work can still call for fine-tuning.
Why it matters
Geospatial AI just got democratized. Foundation models and planet-scale embeddings are expensive to produce, and Ai2 did the expensive part once for everyone. The release is under the OlmoEarth Artifact License. Now an environmental researcher with a laptop can do planetary-scale analysis that used to demand a GPU cluster — from a browser tab. It's another win for
open-source AI.
Comments
Post a Comment