AI2 Turns Its Earth Foundation Model Into a Data Product
OlmoEarth Studio now exports on-demand satellite embeddings, letting researchers skip the heavy lifting of running foundation models themselves
Written by OutOfToken AI
August 12, 2026 · 4 min read · Synthesized from reporting by Hugging Face Blog · How this works
The Allen Institute for AI has released OlmoEarth embeddings, a new export feature inside OlmoEarth Studio that packages the outputs of its remote-sensing foundation models into ready-to-use vector files. Instead of running inference themselves, researchers and organizations can now request custom embeddings for a specific area and time window and get back compact, shareable data. It's a shift from offering a model to offering a product built on top of it.
From Model to Deliverable
OlmoEarth itself is a family of vision transformer models — including a 89-million-parameter Base version and a 308-million-parameter Large version — trained to interpret Sentinel-1, Sentinel-2, and Landsat imagery, including time series data. Both are already published on Hugging Face, where AI2 also hosts tutorials for computing embeddings or fine-tuning the models for specific tasks. The new embeddings feature in OlmoEarth Studio effectively productizes that pipeline, generating vectors on demand rather than requiring users to run the underlying model directly.
What the Embeddings Actually Do
According to AI2's blog post announcing the release, the embeddings act as a fast, cost-effective entry point into the OlmoEarth ecosystem. Locations with similar surface characteristics cluster together in vector space, while differing landscapes land far apart — a property that makes the embeddings useful for tasks like similarity search, segmentation, change detection, and unsupervised exploration of satellite imagery. AI2 says the embeddings have performed well both in its own benchmarking and in independent evaluations, though it hasn't published specific comparative numbers.
"Exported embeddings ship as Cloud-Optimized GeoTIFFs — lightweight, standards-based files built for easy sharing and downstream tooling."
Built for Practical Deployment
The choice of Cloud-Optimized GeoTIFF as the export format is telling. COGs are a widely adopted geospatial standard that plays nicely with existing GIS tools and cloud storage, meaning organizations can plug OlmoEarth embeddings into pipelines they already run rather than adopting a proprietary format. AI2 frames this as part of a broader trend: annual, global, high-resolution embeddings are increasingly treated as reusable infrastructure rather than one-off model outputs, with research groups and startups alike building tooling directly on top of pre-computed vectors instead of raw imagery.
Where OlmoEarth Fits
OlmoEarth models already underpin Skylight, AI2's ocean-intelligence platform, and the broader OlmoEarth Platform is positioned for geospatial inference at what AI2 describes as planetary scale. A more efficient v1.1 model family, announced separately, reportedly cuts compute costs significantly compared to earlier versions, though exact figures for that improvement weren't detailed in the embeddings announcement itself. The embeddings feature effectively sits downstream of that infrastructure, turning heavyweight foundation-model inference into something closer to a self-serve API call.
By exporting embeddings rather than requiring direct model access, AI2 is betting that the real value of Earth-observation AI lies in making foundation-model outputs consumable, not just accurate. If similarity search and segmentation workflows increasingly run on cached vectors instead of raw satellite feeds, OlmoEarth's embedding exports could become a quiet but foundational layer for climate, conservation, and geospatial-analysis teams that have no interest in training or hosting models themselves.
Editorial Note
The research corroborates all major factual claims in the article regarding model parameters, training data sources, supported tasks, and export formats. Sources from AI2's official channels and Hugging Face documentation provide direct confirmation. The article's acknowledgment that AI2 hasn't published specific comparative benchmarking numbers is also accurate per the available sources.
Claim Tracker
AI-assessed
Source 5 (Hugging Face - OlmoEarth-v1-Large) and Source 6 (Hugging Face - OlmoEarth-v1-Base) both confirm these exact parameter counts and model architecture.
Sources 5 and 6 both explicitly state the models are trained on 'Sentinel-1, Sentinel-2, and Landsat images or image time series.'
Source 2 (AI2 blog post) states embeddings 'support a wide range of downstream tasks, from similarity search to segmentation to unsupervised exploration,' and Source 4 mentions 'few-shot segmentation, change detection, similarity search.'
Source 2 (AI2 blog post) confirms 'The exported Cloud-Optimized GeoTIFFs (COGs) are lightweight and easy to share.'
Source 2 states 'OlmoEarth embeddings have shown strong performance in our own benchmarking and in independent evaluations' but provides no specific numbers, confirming the article's claim.
Ask AI about this story
// discussion
sign in to join the discussion