OlmoEarth v1.3: AllenAI's Open Satellite Embeddings Model for Earth
By Ali Al-Montaser — Senior AI Technical Editor
AllenAI launched OlmoEarth v1.3 on Hugging Face, an open satellite embeddings model delivering 128-dimensional vectors per 10-meter pixel globally. It leads 20 of 27 benchmarks against AlphaEarth Foundations and Tessera v2, with training costs of just $2,720 in H100 hours and 53% of v1.2's compute time. Annual global embeddings (~150 TB/year) and a free browser-based Explorer are available now under the OlmoEarth Artifact License.
Executive Overview
AllenAI has released OlmoEarth v1.3 on Hugging Face, an open satellite embeddings model that generates a 128-dimensional vector for every 10-meter pixel on Earth. The release includes ready-to-use annual global embeddings (~150 TB/year) and a free browser-based OlmoEarth Embeddings Explorer requiring no registration. Available immediately via Hugging Face and Source Cooperative under the open OlmoEarth Artifact License, the model leads 20 of 27 benchmarks against AlphaEarth Foundations and Tessera v2, while costing only ~$2,720 in H100 hours to train—53% of v1.2's compute time and 18% of v1's.
📊 Official Data & Technical Specifications Card
| Technical Axis | Confirmed Official Data |
|---|---|
| 💰 Pricing & Usage Cost | Completely free and open source (OlmoEarth Artifact License). Reproduction cost of OlmoEarth v1.3 Base training: ~$2,720 in rented H100 hours for the final run. No token fees or subscriptions. |
| 🌐 Platforms & Immediate Availability | Hugging Face (weights and model), Source Cooperative (ready embeddings ~150 TB/year), OlmoEarth Embeddings Explorer (browser, no registration), OlmoEarth Studio (temporal and resolution customization), source code and technical report publicly available. |
| ⚡ Performance & Speed Metrics | Leads 20 of 27 tests (128-dim version) and 18 of 27 (64-dim version). Compute time: 53% of v1.2 and 18% of v1. Carbon emissions: 50% of v1.2 and 22% of v1. Preliminary analysis in Explorer takes seconds instead of hours of fine-tuning. |
| 🛡️ Security & Breach Resistance | Not applicable (non-generative embedding model). Transparency feature: provenance tables for each 20 km/year tile link every pixel to original Sentinel-1, Sentinel-2, and Landsat scenes with public catalog links. |
| 🧠 Context Window | Not applicable (computer vision model). Embedding dimension: 128 (or compact 64). Inputs: annual time series from Sentinel-1 (radar) and Sentinel-2 and Landsat (multispectral). |
| 🌍 Arabic Language & Regional Support | Global coverage includes land and coastal waters, meaning full coverage of Arab countries. Interface and tool are currently in English, but embeddings and data are available for research and development in the region without linguistic restrictions. |
Deep-Dive Features & Architecture
OlmoEarth v1.3 is built on a Vision Transformer architecture that reads a full year of radar data (Sentinel-1) and multispectral data (Sentinel-2 and Landsat), passes it through an Attention Pooler to aggregate a single vector per land patch, and finally a linear layer compresses it to 128 numbers. The first 64 dimensions operate independently as a compact embedding, enabling flexibility in the size-accuracy trade-off. This "compute once, reuse everywhere" approach turns crop mapping, land cover, and change analysis into a lightweight modeling problem: train a small classifier on top of the embeddings and run it locally without a GPU or downloading satellite imagery.
This release is the cheapest to train in OlmoEarth's history: it consumed 53% of the compute time compared to v1.2 and 18% compared to v1, with 50% and 22% of the carbon emissions, respectively, while maintaining fine-tuning performance equivalent to v1.2 across research labels and partners. The embeddings archive is ~150 TB per year at 10-meter resolution, covering globally 2025, 2024, and 2023, while 2017–2022 covers selected countries with future expansion. Each embedding is accompanied by provenance tables for each 20 km/year tile linking pixels to original satellite scenes.
Benchmark & Competitive Performance
In an evaluation against AlphaEarth Foundations (AEF) and Tessera v2 across 27 tests on six continents, the 128-dimensional version achieved the highest scores in 20 of 27 tests and ranked first overall, while the 64-dimensional version (half the size) led 18 of 27 tests. The team introduced the new PLANTEUR benchmark for tropical crop classification, where the embeddings showed the best generalization to locations unseen by the classifier. In practice, a small classifier suffices to map smallholder crops in Kenya, detect brick kilns in West Dhaka, and assess fire damage in Indonesia, demonstrating the model's real-world utility.
Industry Impact & Enterprise Adoption
OlmoEarth v1.3's open license and free global embeddings lower the barrier to entry for environmental monitoring, agriculture, and disaster response. Organizations can now perform rapid analysis without expensive GPU infrastructure or proprietary data, fostering innovation in climate tech and sustainable development. The model's efficiency—53% of v1.2's compute and 50% of its carbon emissions—also aligns with green AI initiatives, making it attractive for enterprises with sustainability goals. With full coverage of Arab countries, regional researchers and developers can leverage the embeddings for local applications, from desertification tracking to coastal water monitoring, despite the English-only interface.
Conclusion
OlmoEarth v1.3 represents a significant leap in open Earth observation AI, combining state-of-the-art performance with unprecedented accessibility. By providing free, high-resolution embeddings and a user-friendly explorer, AllenAI empowers a global community to derive insights from satellite data without specialized hardware. As the model continues to expand its temporal and geographic coverage, it is poised to become a foundational tool for monitoring our changing planet.
Media Source: Hugging Face | البيان الرسمي للشركة: المصدر الأصلي | Fact Verification & Analysis: AI Tools Oasis
Frequently Asked Questions
Reproducing the training of OlmoEarth v1.3 Base costs approximately $2,720 in rented H100 GPU hours for the final run, according to AllenAI's official statement. The model also consumed 53% of the compute time compared to v1.2 and 18% compared to v1, with 50% and 22% of the carbon emissions, respectively.
Yes, the weights, data, and code are fully open under the OlmoEarth Artifact License, the same open license used in v1 and v1.2. Annual global embeddings (150 TB/year) are available for free on Source Cooperative, and the OlmoEarth Embeddings Explorer runs in the browser without registration or an account.
The model produces a single 128-dimensional vector per 10-meter pixel on Earth's surface, with the first 64 dimensions usable as a standalone compact embedding. It reads a full year of Sentinel-1 radar and Sentinel-2 and Landsat multispectral data, covering land and adjacent coastal waters.
In 27 benchmark tests across six continents, OlmoEarth v1.3's 128-dimensional embeddings achieved the highest scores in 20 of 27 tests and ranked first overall. The 64-dimensional version (half the size) leads 18 of 27 tests, with particular strength in generalizing to unseen locations under the new PLANTEUR benchmark for tropical crops.
The model is available via Hugging Face Models, ready-made embeddings on Source Cooperative, the OlmoEarth Embeddings Explorer in-browser without registration, and OlmoEarth Studio for temporal and resolution customization, plus the technical report and source code. Full global coverage is available for 2025, 2024, and 2023, while 2017–2022 covers selected countries.
AI Tools Oasis Team
Bringing you the latest news and analysis in the world of Artificial Intelligence with accuracy and credibility. Follow us for all updates.
