Tool Reviews

OlmoEarth Version 1.3 provides state-of-the-art embeddings for the entire planet

OlmoEarth Version 1.3 releases global 10-meter satellite embeddings, an explorer, strong benchmarks, lower compute, and open data for mapping.

OlmoEarth Version 1.3 provides state-of-the-art embeddings for the entire planet

Key Points Summary

  • OlmoEarth v1.3 releases planet-scale satellite embeddings at 10-meter resolution in 64 and 128 dimensions, along with a no-account, no-GPU web Explorer.
  • The embeddings compress Sentinel-1, Sentinel-2, and Landsat time-series data into reusable per-pixel vectors, making crop mapping, land-cover classification, and change detection much cheaper.
  • In 27 evaluations, the 128-dimensional embeddings scored best on 20 tasks and ranked highest overall; the 64-dimensional version led on 18 tasks.
  • The model is more efficient and open: it cut training compute and carbon use sharply, with open weights, data, code, and provenance links back to source imagery.
  • The Embeddings Explorer lets users browse years, compare landscapes, find similar locations, train lightweight classifiers from a few labels, and export results as GeoTIFFs.

OlmoEarth v1.3, a new satellite embedding model, is out today, and it arrives with more than weights: precomputed annual embeddings for the entire planet at 10-meter resolution in 64 and 128 dimensions, plus a browser-based tool called the OlmoEarth Embeddings Explorer that lets anyone browse, compare and classify those embeddings without an account or a GPU.

OlmoEarth Version 1.3 provides state-of-the-art embeddings for the entire planet

The release matters because it changes who can work with satellite data. Until now, the main route into OlmoEarth was fine-tuning inside OlmoEarth Studio: training a model per task and running it over imagery, which required GPU access and the ability to download and store the scenes. Embedding models take a different approach, compressing satellite data into reusable features that can serve many tasks cheaply. OlmoEarth v1.3 pushes that idea to its logical end by doing the expensive part once, globally, and handing the results to the community.

One Vector per Pixel, Computed Once

An embedding captures what a model learned from millions of images. OlmoEarth v1.3 reads a year of radar (Sentinel-1) and multispectral (Sentinel-2 and Landsat) time series and outputs a single 128-dimensional vector for every 10-meter pixel on Earth. The first 64 dimensions also stand alone as a smaller embedding.

The pipeline is straightforward: a vision transformer reads the satellite data, an Attention Pooler pools it into one vector per patch of land, and a linear layer reduces it to 128 numbers. Because foundation models and global embeddings are costly to produce, the team computed them for the whole planet and published them on Source Cooperative. From there, 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.

The archive is large. It holds embeddings at 10-meter resolution, roughly 150 TB per year, covering land and adjacent coastal waters. At launch, 2025, 2024 and 2023 are global, while 2017 to 2022 cover select countries, with more coverage promised. Users can pull either the full 128-dimensional version or the lighter 64-dimensional one. Each 20-kilometer tile and year comes with provenance tables listing the Sentinel-1, Sentinel-2 and Landsat scenes the model read, linked to the public catalogs that host them, so any pixel can be traced back to its source imagery.

Benchmarks: 20 of 27 Wins at 128 Dimensions

The team evaluated the embeddings against AlphaEarth Foundations (AEF) and Tessera v2 across 27 evaluations spanning six continents and a wide range of tasks. The 128-dimensional OlmoEarth embeddings scored highest on 20 of 27 evaluations and ranked highest overall. The 64-dimensional version, at half the size, still led on 18 of 27.

OlmoEarth Version 1.3 provides state-of-the-art embeddings for the entire planet

As part of the evaluation, the team introduced PLANTEUR, a new crop-type benchmark for tropical farming, and reported that the embeddings generalize best to places a classifier has never seen. A technical report covers the full results and methodology.

The practical claim is that a small classifier trained on these embeddings is enough to map what is on the ground: classifying smallholder crops in Kenya, detecting brick kilns west of Dhaka, assessing fire damage in Indonesia, or tracking gold mining in the Peruvian Amazon.

Efficiency improved as well. OlmoEarth v1.3 was the cheapest model the team has trained, using 53% of the computing time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions respectively. It also remains a strong fine-tuning model, matching v1.2 across research and partner labels. Reproducing the pretraining of OlmoEarth v1.3 Base would cost roughly $2,720 of rented H100 time for the final run. The weights, pretraining data and pretraining code are all open.

Nine Years in One Place

The release illustrates the embeddings' value with a case from Peru's Madre de Dios region, where illegal gold mining around La Pampa peaked in 2017 and 2018, clearing more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve, according to Amazon Conservation's MAAP monitoring program. In February 2019, Peru's Operation Mercurio shut most of it down. The open question since has been whether it stayed down.

The embeddings offer an answer without any model training. Comparing each pixel's vector between 2017 and 2020 lights up the area southeast of the mining zone where digging was spreading. Running the same comparison from 2020 to 2025 shows the signal shifting: the crackdown is visible in the embeddings. But the second map is not dark. According to MAAP, pockets in the northwest kept operating, and the old mining zone keeps changing as land recovers, part regrowth and part renewed digging. In August 2026, the Associated Press reported, Peru sent navy patrols back into the reserve after prosecutors found mining had crossed into Tambopata itself. The reprieve was brief, and pressure is building again.

OlmoEarth Version 1.3 provides state-of-the-art embeddings for the entire planet

An interactive story opens on this stretch of Madre de Dios and walks through several other examples from the region.

What the Explorer Lets You Do

The OlmoEarth Embeddings Explorer is an interactive analysis tool built from the embeddings themselves, rendered as a false-color map of the planet. Users can pan anywhere, flip between years to see how landscapes changed, and pull up the underlying Sentinel-2 imagery to check the map against reality. Clicking a place finds others that look similar; dropping a few labeled points trains a small classifier on the spot, and the result can be exported as a GeoTIFF. No account and no setup are required.

The speed difference is the point. For many teams, testing an idea against OlmoEarth has meant a fine-tuning run on the order of hours. In the explorer, a first pass takes seconds, and entire regions can be analyzed from a handful of labels without a GPU. When a task does call for fine-tuning, the embeddings help decide whether it deserves the compute and where the model falls short.

Getting Started

The release includes the OlmoEarth v1.3 Base model, the precomputed global embeddings, the OlmoEarth Embeddings Explorer and the technical report. For work that needs a specific time window, OlmoEarth Studio runs the same model over chosen dates, modalities and resolution, and a tutorial shows how to generate embeddings on your own compute. The tools are designed to work together: explore in the Embeddings Explorer, fine-tune in Studio when a task demands it, and take the map from there. Everything is published under the OlmoEarth Artifact License, the same open license used for v1 and v1.2.

The immediate takeaway for readers is that planet-scale satellite analysis no longer starts with a GPU cluster. It can start with a browser tab, a handful of labels and a question about a place you care about.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.