AllenAI has shipped OlmoEarth v1.3, a satellite embedding model that compresses a year of Earth observation data into a single vector for every 10-meter pixel on the planet S¹. Previously, working with OlmoEarth required users to obtain GPU resources and download raw satellite images. The latest version provides precomputed embeddings for the entire globe, accessible via a web browser without any login S¹.
My read: This is the first satellite foundation model I've seen that ships with the entire planet pre-embedded and a browser tool on top. The 20-of-27 benchmark win is impressive on paper, but every number in this release is self-reported, including the PLANTEUR benchmark the team created themselves. I'd want an independent group to reproduce those evals before calling it state-of-the-art. What I find genuinely interesting is the cost trajectory: 53% of v1.2's compute, 18% of v1's. If that holds up, it says something about how quickly satellite model training is maturing.
The model behind the embeddings is a ViT-Base with 128 million parameters, paired with a 14.3-million-parameter Perceiver that handles variable-length time series and a tiny linear layer compressing 768-dimensional tokens down to 128 P². It reads 12 months of radar from Sentinel-1 and multispectral imagery from Sentinel-2 and Landsat, then outputs one 128-dimensional vector per pixel S¹. The first 64 dimensions also work as a standalone, smaller embedding S¹.
Self-reported benchmarks show mixed results
AllenAI evaluated the 128-dimensional embeddings against two competing models, AlphaEarth Foundations and Tessera v2, across 27 tasks spanning six continents S¹. The 128-dim version scored highest on 20 of those 27 evaluations. The 64-dim version led on 18, despite being half the size S¹.
All of those results are self-reported. No independent group has reproduced these evaluations, and the results lack reported error bars S¹. The evaluation also introduced PLANTEUR, a new crop-type benchmark for tropical farming that the same team built S¹. A benchmark created by the model's own developers is a useful signal, not a neutral verdict.
The claim that embeddings generalize best to places a classifier has never seen is the team's own characterization, not independently tested S¹.
Browser access replaces GPU provisioning
The practical shift is access. Before v1.3, the main path into OlmoEarth was fine-tuning inside OlmoEarth Studio, which required GPU access and enough storage to hold satellite imagery S¹. The new release puts precomputed annual embeddings on Source Cooperative and wraps them in the OlmoEarth Embeddings Explorer, a browser tool that needs no account and no machine learning background S¹.
The release highlights several challenging geospatial applications, such as identifying smallholder crops in Kenya, spotting brick kilns near Dhaka, evaluating fire damage in Indonesia, and monitoring gold mining activities in the Peruvian Amazon S¹. Those are illustrative examples from the release, not independently verified deployments.
The team has made the model weights, pretraining data, and code openly accessible S¹. As of January 2025, the GitHub training repository has accumulated 323 stars S¹P³. When applied as a fine-tuning model, v1.3 achieved results on par with v1.2. It also required less computing power than prior versions, utilizing 53% of v1.2's compute and 18% of v1's, while reducing carbon emissions to 50% and 22% of those earlier iterations S¹.
For an agritech startup that wants to classify crop types across a region without provisioning GPUs, the precomputed embeddings on Source Cooperative are the first thing to check. Instead of downloading terabytes of raw imagery and spinning up cloud GPUs, an analyst can pull precomputed vectors directly into a local notebook to train a simple classifier S¹. The model card on Hugging Face lists the architecture and parameter counts P². The training code lives on GitHub under the allenai/olmoearth_pretrain repository P³.
Sources: S1 — OlmoEarth v1.3: State-of-the-art embeddings for the whole planet · P2 — allenai/OlmoEarth-v1_3-Base · Hugging Face · P3 — allenai/olmoearth_pretrain · P4 — stactools-packages/sentinel1 · P5 — ecovision-uzh/ssl4eco
Written from 5 sourced items, 5 of them primary.
