A Free AI Tool Now Lets Anyone Map the Entire Planet, Year by Year

Key points

• OlmoEarth v1.3 outperformed two rival satellite AI systems on 20 of 27 benchmark tests across six continents.
• Pre-built numerical summaries of every 10-metre pixel on Earth are freely available for 2023 through 2025, with partial coverage back to 2017.
• Pre-training the base model cost roughly $2,720 of rented H100 computing time, and all weights, data and code are published openly.
• A browser tool called the OlmoEarth Embeddings Explorer lets anyone train a basic land-cover classifier in seconds from a handful of clicked points, with no sign-in needed.
• Version 1.3 used 53% of the computing time of its predecessor and produced roughly half the carbon emissions.

The Allen Institute for AI's new satellite model can tell you whether a gold mine in the Peruvian Amazon is still operating, whether a crop field in Kenya changed this year, or whether flood damage in Indonesia has cleared. No code required.

The Allen Institute published OlmoEarth v1.3 on Thursday, releasing the model through Hugging Face alongside a browser tool and a library of pre-built data covering the whole planet.

What is an embedding, and why does it matter?

An embedding is a compressed numerical description of something, here a patch of land. Think of it as a very detailed label that a computer can compare and sort. OlmoEarth v1.3 reads a full year of radar and colour satellite data for each 10-metre square of ground and distils it to 128 numbers. Those numbers capture what the land looks like, how it shifts through the seasons, and how it relates to neighbouring patches.

The practical payoff is real. Once those summaries exist, heavy computing power isn't needed to ask questions. Find every field that looks similar to one you already know? Done in seconds. We've tracked how cheaper training approaches are shrinking the barrier for exactly this kind of work, most recently when we reported on a $16 training run that produced six specialised models and a method for boosting a small model with just 100 training steps and a free GPU. OlmoEarth v1.3 pushes the same logic to a planetary scale.

What does the tool actually show?

The Embeddings Explorer opens on Peru's Madre de Dios region, and the reason is instructive. Illegal gold mining around La Pampa cleared more than 1,200 hectares of Amazon forest by 2018. Peru's Operation Mercurio shut most of it down in February 2019. The embeddings make that crackdown visible without any model training: compare each pixel's numerical fingerprint across years and the mining scar lights up, then fades, then shifts.

That story doesn't end cleanly. Amazon Conservation's MAAP program documented continued activity in parts of the zone. In August 2026 the Associated Press reported Peru sent navy patrols back into Tambopata National Reserve after prosecutors found mining had crossed inside the protected boundary. The embeddings register those shifts too.

The Explorer lets any user pan across the globe, flip between years from 2017 to 2025, click a location to find similar-looking places, and export results as a standard map file. No account or setup needed.

Should researchers trust the benchmarks?

The team tested v1.3 against AlphaEarth Foundations and Tessera v2 across 27 tasks on six continents. OlmoEarth's 128-number embeddings ranked first on 20 of those tasks; the lighter 64-number version led on 18. The researchers also released PLANTEUR, a new benchmark for tropical crop classification built to test whether a model generalises to farming regions it has never encountered before.

The efficiency trajectory is the part worth watching. Each generation of OlmoEarth has cost less to train than the last, and v1.3 produced half the carbon emissions of v1.2 while matching its accuracy on fine-tuning tasks. That matters for any organisation building satellite monitoring tools without a large compute budget. For journalists, aid workers or local governments wanting to track land change without hiring a data scientist, the Explorer is the most accessible entry point this kind of technology has offered.

Common questions

Do I need any technical skills to use the Embeddings Explorer?

No. The browser tool requires no account, no installation, no coding. You click places on the map, flip between years, and comparisons happen automatically.

How current is the satellite data?

Global embeddings cover 2023, 2024 and 2025 at launch. Coverage for 2017 through 2022 exists for selected countries, with the team saying more is coming.

Is the model free to use commercially?

The model, its weights and its code are all published under the OlmoEarth Artifact License, the same open licence used for earlier versions. Check the licence text directly before any commercial deployment.