--- title: Introducing OlmoEarth embeddings: Custom embedding... description: OlmoEarth Studio, Ai2's platform for building Earth observation models, now supports computing and exporting embedding vectors from its open-source OlmoEarth... canonical: https://daily.dev/posts/introducing-olmoearth-embeddings-custom-embedding-exports-from-olmoearth-studio-for-downstream-anal-vuymc7ghg twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis | daily.dev og:description: OlmoEarth Studio, Ai2's platform for building Earth observation models, now supports computing and exporting embedding vectors from its open-source OlmoEarth... og:url: https://daily.dev/posts/introducing-olmoearth-embeddings-custom-embedding-exports-from-olmoearth-studio-for-downstream-anal-vuymc7ghg og:image: https://api.daily.dev/og/posts/vuymC7ghg.png og:image:alt: Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis **[Hugging Face](https://daily.dev/sources/huggingface)** · 8 min read · 1 upvotes · 0 comments ## Summary OlmoEarth Studio, Ai2's platform for building Earth observation models, now supports computing and exporting embedding vectors from its open-source OlmoEarth foundation models. Users choose an area of interest, time span, encoder variant (Nano, Tiny, or Base), spatial resolution, and imagery sources (Sentinel-2, Sentinel-1, or both) to get back a Cloud-Optimized GeoTIFF with int8-quantized embedding vectors. The post walks through four applications: similarity search (finding visually similar locations via cosine similarity), few-shot segmentation (training a logistic regression on just 60 labeled pixels to map mangrove vs water vs other with 0.84 weighted F1), change detection (comparing embeddings across time to spot burn scars), and unsupervised PCA visualization to reveal landscape structure. Code, tutorials, and a Colab notebook are provided, and Studio also supports supervised fine-tuning for higher-performance applications. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How are OlmoEarth embedding vectors stored and what format do they come in? OlmoEarth Studio exports embeddings as Cloud-Optimized GeoTIFFs with one band per embedding dimension, stored as signed 8-bit integers (int8) ranging from -127 to +127, with -128 reserved for nodata values. To recover floating-point vectors, the dequantize_embeddings function in the olmoearth_pretrain repository can be used. _Teams evaluating geospatial foundation models can follow release details like this on daily.dev._ ### What encoder variants are available for OlmoEarth embeddings and how do they differ? OlmoEarth Studio offers three encoder variants: Nano (128 dimensions, 1.4M parameters), Tiny (192 dimensions, 6.2M parameters), and Base (768 dimensions, 89M parameters). Tiny is lightweight but still highly performant for tasks like similarity search and few-shot segmentation, while Base encodes richer representations at higher compute and storage cost. _Developers weighing model size against compute cost can track comparisons like this on daily.dev._ ### How few labeled samples are needed to train a land-cover classifier on OlmoEarth embeddings? A logistic regression trained on just 60 labeled pixels (20 per class) over a mangrove region in Ca Mau, Vietnam achieved a weighted F1 score of 0.84 for classifying mangrove, water, and other land types. Accuracy barely improved when increasing to 300 labels, since the pretrained embeddings already capture most of the relevant structure. _Practitioners building few-shot classifiers can find data points like this on daily.dev._ ## Similar posts on daily.dev - [OlmoEarth v1.1: A more efficient family of models](https://daily.dev/posts/olmoearth-v1-1-a-more-efficient-family-of-models-huudwn3vu) · Hugging Face · 0 upvotes · 0 comments - [How to Use AlphaEarth for Similarity Search in Google Earth Engine](https://daily.dev/posts/how-to-use-alphaearth-for-similarity-search-in-google-earth-engine-cmhgssprv) · freeCodeCamp · 2 upvotes · 0 comments - [The OlmoEarth Platform: Geospatial inference at planetary scale](https://daily.dev/posts/the-olmoearth-platform-geospatial-inference-at-planetary-scale-xlouqv9dh) · Hugging Face · 0 upvotes · 0 comments - [Expanding visual search for online art collections](https://daily.dev/posts/expanding-visual-search-for-online-art-collections-1qazg90yx) · Q42 Engineering · 0 upvotes · 0 comments --- Tags: [#ai](https://daily.dev/tags/ai), [#python](https://daily.dev/tags/python), [#embeddings](https://daily.dev/tags/embeddings) [View this post on daily.dev](https://daily.dev/posts/introducing-olmoearth-embeddings-custom-embedding-exports-from-olmoearth-studio-for-downstream-anal-vuymc7ghg) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis","url":"https://daily.dev/posts/introducing-olmoearth-embeddings-custom-embedding-exports-from-olmoearth-studio-for-downstream-anal-vuymc7ghg","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/introducing-olmoearth-embeddings-custom-embedding-exports-from-olmoearth-studio-for-downstream-anal-vuymc7ghg"},"datePublished":"2026-08-12T16:15:31.367Z","dateModified":"2026-09-13T21:09:23.834Z","description":"OlmoEarth Studio, Ai2's platform for building Earth observation models, now supports computing and exporting embedding vectors from its open-source OlmoEarth...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4b0dc7d4930bed4022f1fa9c9c2f7e86?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/4b0dc7d4930bed4022f1fa9c9c2f7e86?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hugging Face","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hugging Face","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/f1f55c67d81a4330acf5b90b26b0c8e1","url":"https://daily.dev/sources/huggingface"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/introducing-olmoearth-embeddings-custom-embedding-exports-from-olmoearth-studio-for-downstream-anal-vuymc7ghg","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,python,embeddings","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hugging Face","item":"https://daily.dev/sources/huggingface"},{"@type":"ListItem","position":3,"name":"Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/introducing-olmoearth-embeddings-custom-embedding-exports-from-olmoearth-studio-for-downstream-anal-vuymc7ghg#faq","mainEntity":[{"@type":"Question","name":"How are OlmoEarth embedding vectors stored and what format do they come in?","acceptedAnswer":{"@type":"Answer","text":"OlmoEarth Studio exports embeddings as Cloud-Optimized GeoTIFFs with one band per embedding dimension, stored as signed 8-bit integers (int8) ranging from -127 to +127, with -128 reserved for nodata values. To recover floating-point vectors, the dequantize_embeddings function in the olmoearth_pretrain repository can be used. Teams evaluating geospatial foundation models can follow release details like this on daily.dev."}},{"@type":"Question","name":"What encoder variants are available for OlmoEarth embeddings and how do they differ?","acceptedAnswer":{"@type":"Answer","text":"OlmoEarth Studio offers three encoder variants: Nano (128 dimensions, 1.4M parameters), Tiny (192 dimensions, 6.2M parameters), and Base (768 dimensions, 89M parameters). Tiny is lightweight but still highly performant for tasks like similarity search and few-shot segmentation, while Base encodes richer representations at higher compute and storage cost. Developers weighing model size against compute cost can track comparisons like this on daily.dev."}},{"@type":"Question","name":"How few labeled samples are needed to train a land-cover classifier on OlmoEarth embeddings?","acceptedAnswer":{"@type":"Answer","text":"A logistic regression trained on just 60 labeled pixels (20 per class) over a mangrove region in Ca Mau, Vietnam achieved a weighted F1 score of 0.84 for classifying mangrove, water, and other land types. Accuracy barely improved when increasing to 300 labels, since the pretrained embeddings already capture most of the relevant structure. Practitioners building few-shot classifiers can find data points like this on daily.dev."}}]} ```