🔍 Read the full analysis: OlmoEarth V1.3 And The Challenge Of Modeling The Whole Planet on ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
The OlmoEarth team released version 1.3 of its satellite embedding model, precomputed annual global data at 10-meter resolution, and a browser-based explorer. The team says its embeddings ranked highest on 20 of 27 evaluations, but those results have not been independently verified in the supplied source material.
The OlmoEarth team has released version 1.3 of its satellite embedding model, along with annual global data at 10-meter resolution and a browser-based explorer. The team says its model ranked highest on 20 of 27 evaluations, but those comparisons are developer-reported and have not been independently verified in the supplied material. The release could make satellite-based land mapping easier to access by letting users work with precomputed data rather than process large image archives themselves.
OlmoEarth v1.3 turns yearly time series from Sentinel-1 radar and multispectral imagery from Sentinel-2 and Landsat into a 128-number representation, or embedding, for each 10-meter pixel. The team also provides a smaller version using the first 64 dimensions. The precomputed global data is available through Source Cooperative, according to the release description.
The accompanying OlmoEarth Embeddings Explorer lets users view the data as false-color maps, switch between years and compare it with underlying Sentinel-2 imagery. The team says users can select a location to find visually similar places or add labeled points to train a small classifier. The explorer is presented as requiring no sign-in or advanced technical knowledge, although the source material does not detail access needs for large-scale analysis.
In comparisons with AlphaEarth Foundations and Tessera v2, the team says the 128-dimensional embeddings ranked highest in 20 of 27 evaluations and the 64-dimensional version in 18. The evaluations covered multiple tasks and six continents. The team also introduced PLANTEUR, a benchmark focused on tropical crop types, and says its embeddings generalized best to locations not seen by a classifier during training. Full results and methods are in the team’s technical report.
Lowering Barriers to Satellite Mapping
Precomputed embeddings offer an alternative to downloading and processing extensive satellite archives for every land-cover or change-mapping project. A researcher can use the vectors as shared features for different tasks, potentially reducing storage and computing demands and making some analyses more practical for teams without dedicated GPU resources. The explorer also offers a way to inspect patterns before building a specialized workflow.
That convenience does not make the resulting maps self-validating. An embedding can help identify places with similar patterns or show that a pixel’s representation changed between years, but it does not by itself explain what caused a change. Applications such as crop monitoring, conservation or land-use enforcement still need suitable labels and checks against imagery or other evidence. The benchmark counts are a useful starting point, not proof that the model will perform equally well in every region or operational setting.
Products Worth Considering
Bose TV Speaker (2nd Gen) – Compact Soundbar with Bluetooth and HDMI‑ARC Connectivity, Clear Dialogue, Dolby Atmos, Deep Bass, Remote Contro
LG UltraGear 32GX850A-B 32-inch 4K UHD OLED Gaming Monitor
Samsung Odyssey OLED G5 27-Inch QHD Gaming Monitor
As an affiliate, we earn on qualifying purchases.
From Task-Specific Models to Reusable Data
Earlier OlmoEarth workflows centered on fine-tuning a model for a particular task in OlmoEarth Studio and then running it over satellite imagery. The team says that approach can require GPU access and downloading and storing imagery. With embeddings, imagery is processed into reusable numerical features that users can apply to different mapping questions or compare across years.
The developers report that v1.3 used 53% of v1.2’s computing time and 18% of the first version’s time to train. They also report emissions equal to 50% of v1.2 and 22% of v1, and say v1.3 matched v1.2 in fine-tuning performance across research and partner labels. The source material does not provide the underlying emissions accounting details. The team estimates that reproducing the final pretraining run for OlmoEarth v1.3 Base would cost about $2,720 in rented H100 computing time; actual costs may vary by provider and setup.
The release describes an example examining landscape change in Peru’s Madre de Dios region, including around La Pampa. The team points to differences between maps for 2017–2020 and 2020–2025. The source cites Amazon Conservation’s MAAP monitoring program and a 2019 crackdown known as Operation Mercurio as relevant history. It also references an Associated Press report dated August 2026 about renewed navy patrols after prosecutors found mining had crossed into Tambopata. Such map differences indicate detected change; attribution to mining or recovery requires other evidence.
“OlmoEarth’s 128-dimensional embeddings score highest on 20 of 27 evals and rank highest overall.”
— The OlmoEarth team
Independent Tests and Data Limits
The supplied material reports the team’s own evaluation results and says the technical report contains the methods, but it does not provide an independent replication of the 27 evaluations. How tasks were selected, how performance varies across regions and how well results carry over to different labels will matter when judging the claims for a particular use.
The release description also does not specify how often the annual global embeddings will be updated, what data gaps or quality limits apply in all locations, or what access and computing requirements users may face for large projects. The reported Peru example does not establish the cause of detected changes on its own. Users would need to check those patterns against imagery, monitoring records or field evidence before drawing conclusions.
Testing the Release in Practice
Researchers and other users can access the precomputed data through Source Cooperative and try the browser explorer, according to the team. The developers point to their technical report for evaluation methods and results. Further testing across regions, tasks and independent datasets will help establish whether the reported benchmark advantage holds outside the team’s own comparisons.
Practical assessment will also depend on the availability and update schedule of the global data, documentation of coverage and quality limits, and validation against reliable reference information. The source material does not announce a date for new datasets or independent evaluations. Until those details emerge, the release establishes what the team has made available and what it reports about performance, while leaving broader reliability and operational fit open to testing.
Key Questions
What is new in OlmoEarth v1.3?
The release includes the v1.3 satellite embedding model, annual global embeddings at 10-meter resolution, a browser-based explorer and a smaller 64-dimensional representation alongside the 128-dimensional version.
What does an embedding represent?
An embedding is a compact set of numerical values learned from imagery. OlmoEarth v1.3 produces a 128-number vector for each 10-meter pixel from yearly satellite time series, which users can reuse for mapping tasks.
Are the reported benchmark results independently confirmed?
Not in the supplied material. The claim that the model ranked highest on 20 of 27 evaluations comes from the OlmoEarth team. The team’s technical report is cited for methods and results, but independent replication is not described.
Can the embeddings identify why a landscape changed?
No. They can indicate that a location’s satellite-derived representation changed between years, but that signal does not establish a cause. Users need to check interpretations against imagery, monitoring data or field evidence.
Primary source: Hugging Face · via ThorstenMeyerAI.com
