OlmoEarth is Ai2's family of open models for satellite images; version 1.3 turns a year of images of each spot on Earth into a short list of numbers, called an embedding.
In plain words
For every 10-metre patch of land, OlmoEarth v1.3 takes a year of radar and multispectral images from the Sentinel-1, Sentinel-2 and Landsat satellites and outputs 128 numbers. Ai2 has computed these for the whole planet each year, so a crop map or change study needs only a small classifier on top. Ai2 says they scored highest on 20 of 27 tests against two rival sets, and it has released the weights, training data and code.
A quick example
Ai2 shows how comparing each patch's embedding across 2017, 2020 and 2025 reveals where illegal gold mining in Peru spread and later slowed, with no model training needed.
Where you'll meet it
- Crop, land-cover and change maps.
- Finding places that look similar on satellite images.
- Tracking things like mining or fire damage over time.
Related
Sources
- DocsOlmoEarth v1.3: State-of-the-art embeddings for the whole planet, Ai2 (Hugging Face blog) · read 22 Oct 2026
- Docsallenai/OlmoEarth-v1_3-Base, Ai2 (Hugging Face) · read 22 Oct 2026
Checked by WhatIsAI against the sources above · 22 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.