AivexaNewsSearch
AI news for builders and product teamsChecked every hour

OlmoEarth v1.3: State-of-the-art embeddings for the whole planet

Collected Oct 22, 2026

Hugging Face released OlmoEarth v1.3, a satellite embedding model, together with precomputed annual embeddings covering the whole planet at 10-meter resolution, available in 64- and 128-dimensional versions. The release also includes a browser-based tool, the OlmoEarth Embeddings Explorer, and a technical report. Model weights, pretraining data and pretraining code are published openly under the OlmoEarth Artifact License, the same license used for the earlier v1 and v1.2 models.

The core idea is to compute once and reuse everywhere. The model reads a year of radar imagery from Sentinel-1 and multispectral imagery from Sentinel-2 and Landsat, then emits one 128-number vector for every 10-meter pixel. The first 64 of those dimensions stand alone as a smaller embedding. Architecturally, a vision transformer reads the satellite time series, an attention pooler condenses it into a single vector per patch of land, and a linear layer reduces that to 128 numbers. Because the vectors are precomputed and shared on Source Cooperative, downstream work such as crop mapping, land-cover classification or change analysis becomes a small modeling job: train a lightweight classifier on top and run it locally.

That is a meaningful change from the previous workflow. Until now the main route was fine-tuning in OlmoEarth Studio, which meant training a task-specific model and running it over imagery — requiring GPU access plus the ability to download and store the source data. Embeddings shift the expensive part to a one-time foundation-model run and let many tasks share the output. In practice, a user who wants a first-pass answer no longer needs a training pipeline at all.

The evaluation compared OlmoEarth v1.3 embeddings with AlphaEarth Foundations (AEF) and Tessera v2 across 27 assessments spanning six continents and a range of tasks. The 128-dimensional embeddings scored highest on 20 of the 27 and ranked highest overall; the 64-dimensional version, at half the size, led on 18 of 27. The team also introduced PLANTEUR, a new crop-type benchmark for tropical farming, and reports that the embeddings generalize best to locations a classifier has never seen. Concrete applications cited include classifying smallholder crops in Kenya, detecting brick kilns west of Dhaka, assessing fire damage in Indonesia and tracking gold mining in the Peruvian Amazon.

Training efficiency improved too. OlmoEarth v1.3 took 53% of the compute time of v1.2 and 18% of v1, with 50% and 22% of the carbon emissions respectively, making it the cheapest OlmoEarth model to train so far. It also remains a strong fine-tuning model, matching v1.2 on research and partner labels. For anyone reproducing the work, pretraining the Base model would cost roughly $2,720 of rented H100 time for the final run.

A worked example shows what change detection looks like without training. In Peru's Madre de Dios region, illegal gold mining around La Pampa cleared more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve, peaking in 2017 and 2018, according to Amazon Conservation's MAAP monitoring program; Peru's Operation Mercurio shut most of it down in February 2019. Comparing each pixel's vector between 2017 and 2020 highlights expansion to the southeast; comparing 2020 with 2025 shifts the signal, reflecting both recovery and renewed activity. Per MAAP, pockets in the northwest kept operating. The Associated Press reported in August 2026 that Peru sent navy patrols back into the reserve after prosecutors found mining had crossed into Tambopata itself. This suggests embedding deltas can serve as a cheap screening layer for monitoring, though interpreting what a change means still requires ground knowledge.

The Embeddings Explorer renders the archive as a false-color map. Users can pan, switch between years, overlay Sentinel-2 imagery for verification, click a location to find visually similar places, or drop a few labeled points to train a small classifier on the spot and export the result as a GeoTIFF. No account or setup is required. A first pass that might previously have meant hours of fine-tuning takes seconds. My reading is that the tool fits best as a triage step — deciding whether a task justifies fine-tuning and where the model struggles — rather than a replacement for it.

The archive holds about 150 TB per year at 10-meter resolution, covering land and adjacent coastal waters. At launch, 2025, 2024 and 2023 are global, while 2017 through 2022 cover select countries, with more coverage planned. Each 20-kilometer tile and year ships with provenance tables listing the Sentinel-1, Sentinel-2 and Landsat scenes the model read, linked to the public catalogs hosting them, so any pixel can be traced back to its source imagery. For work needing a specific time window, modality or resolution, OlmoEarth Studio runs the same model on user-chosen inputs, and a tutorial covers generating embeddings on your own compute.

Why it matters: teams that previously needed GPUs and imagery storage to experiment with satellite ML can now test ideas in a browser and retrieve ready-made features for their own pipelines. Earth-observation startups, agricultural and conservation groups, and researchers without heavy infrastructure are the clearest beneficiaries. A likely trade-off, in my inference, is that precomputed annual embeddings are fixed to a resolution and cadence, so fast-moving or sub-annual monitoring may still require running the model yourself through Studio.

Read at Hugging Face Blog

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt