OlmoEarth v1.3: Free Satellite Embeddings for the Planet

Satellites photograph the entire planet every few days, but turning that flood of imagery into a useful map has usually required GPUs, storage and machine learning expertise. On October 22, 2026, the Allen Institute for AI (Ai2) released OlmoEarth v1.3, a satellite embedding model, together with precomputed yearly embeddings for the whole planet and a free browser tool for exploring them. Ai2 says anyone can start using it right away, with no sign-in and no advanced technical knowledge.

• Resolution: one 128-dimensional vector for every 10-meter pixel on Earth; the first 64 dimensions also work as a smaller embedding
• Coverage: global for 2023, 2024 and 2025, with select countries for 2017 to 2022, at roughly 150 TB per year
• Benchmarks: the 128-dim version scored highest on 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, per Ai2
• Cost: Ai2 estimates the final pretraining run at roughly $2,720 of rented H100 time

What Are Satellite Embeddings?

An embedding compresses what a model has learned into a compact list of numbers. OlmoEarth v1.3 reads a full year of radar data from Sentinel-1 and multispectral imagery from Sentinel-2 and Landsat, then summarizes each 10-meter patch of land as a single vector. Similar places end up with similar vectors.

That changes the economics of Earth observation. Instead of fine-tuning a large model for every new mapping task, you can train a small classifier on top of the embeddings and run it locally. Ai2 describes examples such as classifying smallholder crops in Kenya, spotting brick kilns near Dhaka and assessing fire damage in Indonesia. The embeddings are hosted on Source Cooperative, and every 20-kilometer tile comes with provenance tables listing exactly which satellite scenes the model read.

How Good Is OlmoEarth v1.3?

Ai2 compared the new embeddings with two other planetary embedding sets, AlphaEarth Foundations and Tessera v2, across 27 evaluations on six continents. According to Ai2, the 128-dimensional embeddings ranked highest overall and led on 20 of 27 tests, while the half-size 64-dimensional version still led on 18. As part of the work, Ai2 introduced PLANTEUR, a new crop-type benchmark focused on tropical farming, where it says its embeddings generalized best to places a classifier had never seen.

The model is also lighter to build. Ai2 reports that v1.3 needed 53% of the compute time of v1.2 and 18% of the original v1, with 50% and 22% of the carbon emissions respectively, while matching v1.2 on fine-tuning tasks.

How Can I Try the OlmoEarth Embeddings Explorer?

The new OlmoEarth Embeddings Explorer runs in a web browser and renders the embeddings as a false-color map of the planet. You can pan anywhere, flip between years to watch landscapes change, and pull up the underlying Sentinel-2 imagery to check the map against reality. Click a spot to find similar places, or drop a few labeled points to train a small classifier on the spot and export the result as a GeoTIFF. Ai2 says a first pass takes seconds and needs no GPU or account.

For bigger projects, OlmoEarth Studio runs the same model over custom dates and resolutions. The weights, pretraining data and code are all published under the OlmoEarth Artifact License, the same open license Ai2 used for earlier versions.

Why This Release Matters for Open AI Research

Open models usually mean open weights. OlmoEarth goes further by sharing the expensive part, the planet-scale inference, so researchers, conservation groups and agricultural planners can skip straight to their questions. It builds on the platform we covered when OlmoEarth cut wildfire risk mapping to 30 hours. Find more open-model news in our AI section.

Sources: Hugging Face Blog: OlmoEarth v1.3 — October 22, 2026; Ai2: OlmoEarth embeddings — background on earlier releases.