The shift from fine-tuning to embeddings
Allen Institute for AI just released OlmoEarth v1.3, and the real story isn't just the model—it's the 150 TB of precomputed embeddings they computed for the entire planet and made freely available. Until now, using OlmoEarth meant fine-tuning in OlmoEarth Studio: training a task-specific model, running it over satellite imagery, waiting for GPU time. That workflow works, but it's a barrier.
With v1.3, AI2 precomputed annual embeddings at 10-meter resolution for the whole planet (64- and 128-dimensional versions), hosted them on Source Cooperative, and shipped a browser-based explorer that lets you train lightweight classifiers on top—locally, no account, no GPU. The compute-intensive part is done. What's left is a lightweight modeling problem: drop a few labeled points, train a small classifier in seconds, export a GeoTIFF.
This is the geospatial foundation model playbook working the way it's supposed to: compute once, reuse everywhere.
State-of-the-art performance, getting cheaper
OlmoEarth v1.3 reads a year of Sentinel-1 radar plus Sentinel-2 and Landsat multispectral time series and compresses it into one 128-dimensional vector per 10-meter pixel. The first 64 dimensions also work standalone as a smaller embedding. AI2 evaluated these embeddings against AlphaEarth Foundations (AEF) and Tessera v2 across 27 benchmarks spanning six continents.
The 128-dimensional version scored highest on 20 of 27 evaluations and ranked first overall. The 64-dimensional version—at half the size—still led on 18 of 27. Everything behind the model is open: weights, pretraining data, pretraining code.
As part of this evaluation, AI2 introduced PLANTEUR, a new crop-type benchmark for tropical farming. The embeddings generalized best to places the classifier had never seen—exactly the test that matters when you're trying to map smallholder crops in Kenya or track gold mining in Peru.
Training cost trajectory
The cost curve is fascinating. OlmoEarth v1 (the original) would cost roughly $15,060 in rented H100 time to reproduce. v1.2 dropped that to $5,100. v1.3 costs about $2,720—53% of v1.2's compute time and 18% of v1's, with corresponding drops in carbon emissions (50% and 22% respectively).
That's not just incremental improvement—that's the efficiency work you need to make these models economically viable at scale. And v1.3 still matches v1.2's fine-tuning performance across research and partner labels, so you're not trading capability for cost.
Tracking illegal gold mining with embedding distance
The embeddings unlock a new analysis mode: compare each pixel's vector across years, no classifier required. AI2 demonstrates this with La Pampa, Peru, where illegal gold mining peaked in 2017-2018, clearing over 1,200 hectares near the Tambopata National Reserve before Peru's Operation Mercurio shut it down in February 2019.
Compare embeddings from 2017 to 2020, and you see mining spreading southeast. Run the same comparison from 2020 to 2025, and the signal shifts: the crackdown is visible in the embedding change patterns. The old mining zone keeps changing as land recovers, but the map isn't dark—pockets in the northwest kept operating, and MAAP monitoring data confirms the pressure is building again. By August 2026, AP reported Peru sent navy patrols back into the reserve after finding mining had crossed into Tambopata itself.
The embedding distance is doing real analytical work here—no labels, no training, just vector arithmetic revealing landscape change patterns that align with ground truth from investigative journalism and conservation monitoring.
The Embeddings Explorer: seconds, not hours
The OlmoEarth Embeddings Explorer is a browser-based tool that renders the embeddings as a false-color global map. Pan anywhere, flip between years, pull up Sentinel-2 imagery to verify. Click a place to find similar locations. Drop a few labeled points and train a classifier on the spot, then export results as GeoTIFF.
For teams that previously needed a fine-tuning run (hours of GPU time), first-pass analysis now takes seconds. You can test whether a task is worth the fine-tuning compute at all, identify edge cases, or get results directly if the lightweight classifier suffices.
The archive holds embeddings at 10-meter resolution—about 150 TB per year. At launch, 2025, 2024, and 2023 are global; 2017-2022 cover select countries, with more coming. Each embedding includes provenance: for every 20-kilometer tile and year, you get tables listing the Sentinel-1, Sentinel-2, and Landsat scenes the model read, with links to public catalogs so you can trace any pixel back to source imagery.
Why this matters: the cost structure of geospatial AI
The traditional computer vision story is: train a foundation model, release weights, let users fine-tune. That works when inference is cheap and data is abundant. Geospatial doesn't fit that pattern. Satellite imagery is petabyte-scale, fine-tuning requires GPU access and storage infrastructure, and most potential users (conservation groups, smallholder agriculture researchers, municipal planners) don't have those resources.
Precomputed embeddings change the economics. The expensive part—reading millions of satellite time series through a vision transformer—is done once. What remains is a problem you can solve on a laptop: train a small classifier on vectors, run it locally. That's the unlock.
The Embeddings Explorer makes this concrete: you can go from question to prototype in a browser session, no pipeline to set up. If the lightweight approach works, you're done. If you need fine-tuning for better accuracy, you've validated that the compute investment is justified and identified where the embeddings fall short.
What comes next
This release includes the v1.3 Base model, the precomputed global embeddings, the Embeddings Explorer, and a technical report. If you need a specific time window or custom dates, OlmoEarth Studio runs the same model over your chosen parameters, and there's a tutorial for generating embeddings on your own infrastructure. Everything is published under the OlmoEarth Artifact License, the same open terms as v1 and v1.2.
The more interesting question is what happens when this pattern generalizes. If precomputed embeddings plus a lightweight explorer is the right distribution strategy for geospatial foundation models, what other domains have similar cost structures? Where else are we asking users to pay the compute and storage tax for inference when we could precompute representations and ship vectors instead?
AI2 isn't just releasing a model here—they're demonstrating a deployment strategy that makes foundation models accessible to users who couldn't otherwise use them. That's worth paying attention to.