On October 22, 2026 (Beijing time), the Allen Institute for AI (Ai2) released OlmoEarth v1.3. The release has three parts: a foundation model for satellite time series; annual 10-meter embeddings for land and nearby coastal waters (64- or 128-dimensional); and a browser Embeddings Explorer. In the Explorer you can pan the false-color map, flip years, find similar places, label a few points to train a small classifier, and export a GeoTIFF—no account, and no GPU of your own on that path. We have not tested the Explorer yet; if you have, feedback to the site is welcome.
That breaks into three jobs. Comparing each pixel’s embedding across years, or clicking a place to find similar spots, needs no training. Labeling a few points in the Explorer trains a small classifier on frozen embeddings; it does not retrain the foundation model. Full fine-tuning over chosen dates, modalities, and resolution still goes through OlmoEarth Studio, which needs compute and imagery workflows. The Explorer helps you decide whether that compute is worth it; it is not a Studio substitute.
What you see on the map is an annual summary, not a live satellite feed. OlmoEarth v1.3 compresses a year of radar (Sentinel-1) and multispectral (Sentinel-2, Landsat) data into one vector per 10-meter pixel. Ai2 and Source Cooperative say 2023–2025 are global, while 2017–2022 cover select countries and are still being backfilled—a given city may not have a full nine-year stack. Magenta-heavy change maps show where embeddings moved a lot; they do not by themselves say whether the cause was construction, disaster, crops, or clouds. Official demos still open Sentinel-2 and cite external monitoring when they explain what changed.
Ai2 reports that its 128-dimensional embeddings scored highest on 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, with the 64-dimensional version leading on 18 of 27. Those numbers come from Ai2’s own evaluations and are for reference only. Weights, pretraining data, and code are public; precomputed embeddings live on Source Cooperative with provenance tables that link tiles back to public Sentinel-1, Sentinel-2, and Landsat catalogs.
Everything in this release uses the OlmoEarth Artifact License—not Apache. You may use, modify, and distribute with attribution and pass-down of restrictions, but military and defense-related uses are barred, as are extractive uses that plan or facilitate oil, gas, and mineral extraction, including drilling, mining, and deforestation. The Hugging Face model card marks the license as other and points to that agreement. Read the full license before you depend on these artifacts. For another DigIntoAI guide that treats license limits as a selection filter, see FastH3 Omni on a Local GPU — Thresholds, Speed Caveats, License Limits. Nearby open-research releases on Hugging Face include Hugging Face Carbon-X Is an Open DNA Language Model That Runs on One GPU. For open vision-language weights with an explicit local-GPU framing, compare Nemotron 3.5 Super VL and Bolt: Specs, License, Runs on 24GB.