What is OlmoEarth?
OlmoEarth is a domain‑specific inference stack built on top of the OLMo language‑and‑vision architecture. While the original OLMo models focused on text and general image understanding, OlmoEarth ships with a fine‑tuned vision backbone capable of ingesting multi‑spectral satellite rasters, LiDAR point clouds, and temporal climate grids. The model weights are released under the Apache 2.0 license, enabling commercial use, customization, and local deployment.
Unlike traditional geospatial platforms that force users into a specific programming environment (JavaScript in Google Earth Engine) or a closed‑source processing pipeline, OlmoEarth delivers raw AI inference as a building block. Developers can host the model on Hugging Face’s Inference Endpoints, deploy it via Serverless Inference API, or pull the weights into any PyTorch environment. A lightweight JavaScript wrapper (Transformers.js) even enables in‑browser analysis of smaller geo-tiles, a feature particularly useful for field researchers with limited connectivity.
How it works under the hood
At its core, OlmoEarth leverages a vision‑transformer encoder pre‑trained on millions of open‑access satellite scenes and then fine‑tuned on specific downstream tasks such as land‑cover classification, crop‑yield prediction, flood mapping, and urban expansion tracking. The model outputs dense segmentation masks, regression heatmaps, or classification labels, and can be chained with a large language model for natural‑language querying of geospatial results — for example, “Identify deforestation hotspots larger than 5 ha in this Sentinel‑2 tile from Q2 2025.”
The infrastructure side relies on Hugging Face’s distributed inference backend. For batch processing of whole countries or continents, OlmoEarth can be launched on GPU‑accelerated clusters through the standard Inference Endpoints interface. Input data can be streamed directly from cloud object stores (AWS S3, Google Cloud Storage) or from Copernicus Open Access Hub, making it straightforward to connect with existing European Earth‑observation data pipelines.
Pricing and EU availability
Hugging Face offers a free tier via Spaces for quick experimentation, where OlmoEarth can process individual scenes without charge. For production workloads, the platform uses the same transparent pricing as other Inference Endpoints:
- CPU instance: $0.06 / hour (≈ €0.055 / hour)
- GPU instance (T4): $0.60 / hour (≈ €0.55 / hour)
- Larger GPU instances (A10G, A100) scale from $1.50 to $4.00 / hour (≈ €1.38 – €3.68 / hour)
EUR conversions use an indicative rate of 1 USD = 0.92 EUR, as of late July 2026. Crucially, because the model weights are open, you are never forced to pay per inference. Organisations that handle sensitive geospatial data — defence, urban planning, critical infrastructure monitoring — can download the model once and run it on their own hardware, keeping all data inside their perimeter.
Hugging Face’s Inference Endpoints can be deployed in European AWS regions (e.g., eu‑west‑1 in Ireland, eu‑central‑1 in Frankfurt), ensuring that data processing stays within the EU/EEA. The downloaded models also qualify for on‑premise operation, a major advantage for public agencies bound by strict data‑sovereignty rules.
Comparison with existing platforms
To understand where OlmoEarth fits, we compared it with the two dominant players in the planetary‑scale geospatial space. The table below highlights the differences that matter most to European users.
| Platform | Model Openness | Pricing (July 2026) | EU Data Hosting | Primary Interface |
|---|---|---|---|---|
| Google Earth Engine | Proprietary; no direct model access | Usage‑based (Cloud credits); limited free tier | Yes (Google Cloud EU regions) | JavaScript, Python (ee) |
| Microsoft Planetary Computer | Data open; some open models (TorchGeo) | Free catalog access; compute via Azure | Yes (Azure EU regions) | Python (planetary‑computer), REST |
| OlmoEarth (HF + AI2) | Fully open (Apache 2.0) | Free demo; from €0.055/h or self‑hosted | Yes (HF EU endpoints + on‑premise) | REST, Python, JS (Transformers.js) |
Prices are approximate; EUR values calculated at 1 USD = 0.92 EUR. Always check Hugging Face pricing for current rates.
What it means for Europe: AI Act, GDPR, and data sovereignty
Geospatial AI often touches on domains that the EU considers high‑risk under the AI Act — environmental monitoring, critical infrastructure planning, emergency response. Providers of such AI systems must ensure transparency, human oversight, and robust documentation. Because OlmoEarth is fully open‑source, it naturally supports transparency obligations: every architectural detail and training dataset is accessible, allowing conformity assessments that would be impossible with black‑box commercial APIs.
For GDPR, the key benefit is local deployment. A municipality or regional environmental agency can download the model weights and run inference entirely within its own secure cloud tenant or on‑premise server, ensuring that satellite imagery containing indirectly personal data (e.g., high‑resolution imagery of residential areas) never leaves the controlled environment. Hugging Face’s official inference endpoints also offer EU‑only processing, but the option to go fully self‑hosted is a game‑changer for public bodies with the most stringent data‑protection requirements.
Our take
OlmoEarth signals a maturing of the Hugging Face ecosystem: it is no longer just a hub for language models but a genuine multi‑modal infrastructure layer. For European developers and Earth‑observation startups, the platform slashes the barrier between having an idea and running it at scale. The combination of open weights, predictable pricing, and EU‑compliant deployment paths makes it a credible alternative to Google’s and Microsoft’s walled gardens.
We will keep an eye on community benchmarks and plan to run our own latency tests once the platform stabilizes. In the meantime, you can explore the demo on Hugging Face Spaces and check the documentation linked in the original announcement.
Can I use OlmoEarth with my own satellite imagery stored in an S3 bucket in Europe?
Yes. The Inference Endpoint can stream data directly from any S3‑compatible object store, including buckets in the eu‑central‑1 or eu‑west‑1 regions. If you prefer not to move data at all, you can simply download the open model weights and run inference on a machine co‑located with your storage.
How does OlmoEarth’s performance compare to traditional satellite image analysis pipelines?
Early community reports suggest that its vision transformer achieves state‑of‑the‑art accuracy on standard land‑cover benchmarks (over 85 % mIoU) while processing a full Sentinel‑2 scene in under 30 seconds on a single A10G GPU. The real advantage, however, is the elimination of manual feature engineering and the ability to query results in natural language via the integrated LLM connector.
Is OlmoEarth compliant with the EU AI Act for use in public infrastructure projects?
While no single model is inherently “AI Act compliant”, OlmoEarth’s open‑source nature and detailed model cards greatly simplify the conformity assessment process. The model’s training data, architecture, and intended use are fully documented, meeting the transparency requirements for high‑risk AI systems. Public bodies must still conduct their own risk assessment, but the open license removes the vendor‑lock‑in that often hinders auditing.
