Mapping crops, forests, or land cover from satellite imagery has usually meant gathering large volumes of data, training a model for each task, and waiting for GPU time. That approach is powerful, but it can feel heavy when a team simply wants to test an idea.
OlmoEarth v1.3 offers a lighter route. Ai2 has released a satellite embedding model along with precomputed annual embeddings for the entire planet, so a first pass at a mapping question can take seconds.
This guide explains what the release includes, how embeddings work, how to read the benchmark results Ai2 published, and how to put the embeddings to use. Details come from Ai2's release post, and the OlmoEarth overview has more background.
1. OlmoEarth v1.3 at a Glance
OlmoEarth v1.3 is a state of the art satellite embedding model from Ai2. It reads a year of satellite time series and produces one compact description, called an embedding, for every 10 meter pixel on Earth.
| Detail | What is reported |
|---|---|
| Model | OlmoEarth v1.3 Base |
| Output | One 128 dimensional vector per 10 meter pixel |
| Smaller option | The first 64 dimensions, usable on their own |
| Input data | A year of Sentinel-1 radar, Sentinel-2 multispectral, and Landsat multispectral time series |
| Coverage | 2023, 2024, and 2025 global at launch, with 2017 to 2022 for select countries |
| Archive size | Roughly 150 TB per year, covering land and adjacent coastal waters |
| Hosting | Precomputed embeddings on Source Cooperative |
| Tools | OlmoEarth Embeddings Explorer in the browser, plus a technical report |
| License | OlmoEarth Artifact License |
2. What Is a Satellite Embedding?
A satellite embedding is a set of numbers that summarizes what a model has learned about a place from its satellite imagery. Embedding models compress satellite data into ready to use features that can be reused cheaply across many tasks.
A helpful way to picture it is as a compact profile of each patch of land. Instead of working with raw imagery, a simple tool can compare, group, or label these profiles directly. An embedding captures what the model learned from millions of images, so that knowledge is available without repeating the heavy work.
2.1 How is an embedding different from fine tuning?
Until now, the main way to use OlmoEarth was fine tuning in OlmoEarth Studio. That meant training a model for each task and running it over satellite imagery to produce a map. It also required GPU access and the ability to download and store the imagery.
| Approach | What it involves | Where it fits |
|---|---|---|
| Embeddings | Train a small classifier on precomputed features and run it locally | Quick first passes and analysis of entire regions from a handful of labels |
| Fine tuning in Studio | Train a model per task over satellite imagery, with GPU access and imagery storage | Tasks that need a specific time window, modalities, or resolution |
2.2 Why do embeddings matter?
- Lower starting cost: foundation models and global embeddings are expensive to produce, so Ai2 computed them once for the planet and shared them
- Reuse across tasks: a crop map, a land cover map, or a change analysis becomes a lightweight modeling problem
- Local analysis: a small classifier on top of the embeddings can run on your own machine
- Better decisions: embeddings can help a team decide whether a task deserves fine tuning compute, and where the model falls short
3. How Does OlmoEarth v1.3 Create an Embedding?
The short answer: the model reads the satellite data, condenses each patch of land into one vector, and shortens that vector to a fixed size.
| Step | What happens | Why it matters |
|---|---|---|
| Vision transformer | Reads the satellite time series | Learns patterns from radar and multispectral data |
| Attention Pooler | Pools the result into one vector per patch of land | Gives each location a single summary |
| Linear layer | Reduces that vector to 128 numbers | Keeps the embedding small and consistent |
The result is a consistent set of numbers for every location, which makes comparison across places and years straightforward.
3.1 What do 64 and 128 dimensions mean?
Dimensions are the numbers in each embedding. More dimensions can hold more detail, while fewer dimensions are smaller to store and move around.
| Version | Size | Reported result |
|---|---|---|
| 128 dimensions | Full embedding | Highest score on 20 of 27 evaluations, and highest overall rank |
| 64 dimensions | Half the size | Still leads on 18 of 27 evaluations |
4. How Well Do the Embeddings Perform?
Ai2 evaluated OlmoEarth v1.3 against AlphaEarth Foundations and Tessera v2 across 27 evaluations covering six continents and many task types. The embeddings also generalize best to places a classifier has never seen.
As part of this work, Ai2 introduced PLANTEUR, a new crop type benchmark for tropical farming. The model weights, pretraining data, and pretraining code are all open, which allows others to inspect and reproduce the work. Full results and methodology are in the technical report.
These figures are published by Ai2, so independent evaluations and your own tests on a real area give a fuller picture.
4.1 What can a small classifier map?
A small classifier trained on these embeddings can map real world conditions. Examples from the release include:
| Place | What was mapped |
|---|---|
| Kenya | Smallholder crops |
| West of Dhaka | Brick kilns |
| Indonesia | Fire damage |
| Peruvian Amazon | Gold mining |
5. How Efficient Is OlmoEarth v1.3 to Train?
OlmoEarth v1.3 is the cheapest OlmoEarth model to train so far. It used 53% of the computing time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions.
Pretraining the Base model would cost roughly $2,720 in rented H100 time for the final run.
| Version | Base model pretraining cost (rented H100, final run) |
|---|---|
| OlmoEarth v1 | $15,060 |
| OlmoEarth v1.2 | $5,100 |
| OlmoEarth v1.3 | $2,720 |
The model also remains a strong fine tuning model, matching v1.2 performance across research and partner labels. Lower cost matters because it makes the work easier for others to reproduce and build on.
6. What Can Embeddings Reveal Over Time? A Look at La Pampa, Peru
Change analysis is one of the clearest uses of annual embeddings. The release highlights Madre de Dios, Peru, where illegal gold mining around La Pampa peaked in 2017 and 2018. According to Amazon Conservation's MAAP monitoring program, it cleared more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve. In February 2019, Peru's Operation Mercurio shut most of it down.
The open question has been whether it stayed down. Embeddings offer a simple way to look. You compare each pixel's vector across years, with no model training required.
| Period | What the comparison shows |
|---|---|
| 2017 to 2020 | Mining spreading to the southeast |
| 2020 to 2025 | That corner has gone quiet, and the crackdown is visible |
The second map is still not dark. According to MAAP, pockets in the northwest kept operating, and the older mining zone keeps changing through regrowth and renewed digging. The embeddings register that too. In August 2026, the Associated Press reported that Peru sent navy patrols back into the reserve after prosecutors found mining had crossed into Tambopata itself.
7. What Is the OlmoEarth Embeddings Explorer?
The OlmoEarth Embeddings Explorer is an interactive analysis tool built from the embeddings themselves. It renders them as a false color map of the planet, and no account or setup is required.
| Feature | What you can do |
|---|---|
| Pan and flip years | See how landscapes changed over time |
| Sentinel-2 view | Check the map against the underlying imagery |
| Similarity search | Click a place to find others that look similar |
| Quick classifier | Drop a few labeled points, train on the spot, and export a GeoTIFF |
Testing an idea against OlmoEarth has often meant a fine tuning run that takes hours. In the Explorer, a first pass takes seconds. Teams can analyze entire regions from a handful of labels without a GPU.
7.1 How can you run a first analysis?
A practical flow follows directly from those features:
- Look: pan to a place you care about and compare years
- Verify: pull up the Sentinel-2 imagery to confirm what the map suggests
- Search: click a location to find similar places
- Classify: add a few labeled points, train the classifier, and export a GeoTIFF
8. What Coverage and Transparency Does the Archive Offer?
The archive holds embeddings at 10 meter resolution, roughly 150 TB per year, covering land and adjacent coastal waters. Both the 128 dimensional and the lighter 64 dimensional versions are available.
| Years | Coverage at launch |
|---|---|
| 2023, 2024, 2025 | Global |
| 2017 to 2022 | Select countries, with more coverage on the way |
Every embedding also comes with a record of where it came from. For each 20 kilometer tile and year, open provenance tables list the Sentinel-1, Sentinel-2, and Landsat scenes the model read, with links to the public catalogs that host them. This lets you trace any pixel back to the imagery behind it, which supports checking and trust in the results.
9. How Do You Get Started With OlmoEarth v1.3?
The tools are designed to work together. Explore in the Embeddings Explorer, fine tune in OlmoEarth Studio when a task calls for it, and take your map from there.
| Need | Practical starting point |
|---|---|
| A quick first look at a region | The Embeddings Explorer in the browser |
| A custom map from a few labels | A small classifier on the precomputed embeddings |
| A specific time window, modalities, or resolution | OlmoEarth Studio, which runs the same model over your chosen settings |
| Embeddings on your own compute | The tutorial Ai2 provides for generating them yourself |
10. License and Access Notes
Everything is published under the OlmoEarth Artifact License, the same license used for v1 and v1.2. The model weights, pretraining data, and pretraining code are open, and the pretraining code is on GitHub. Read the license text before using the embeddings in a product, since terms matter for commercial use.
11. Conclusion
OlmoEarth v1.3 shifts the first step of satellite mapping from heavy training runs to reusable, precomputed features. Annual 10 meter embeddings for the planet, an Explorer that works in the browser, and open weights, data, and code give teams several practical ways to ask questions of the same model.
The key takeaways are straightforward. Embeddings are compact summaries that can be reused across many tasks. A small classifier on top of them can map crops, land cover, and change. Fine tuning remains available when a task needs more. In Ai2's published benchmarks, the embeddings lead most evaluations, so the most reliable next step is to pull them for a place you care about and test them on your own question.