Ai2 releases OlmoEarth v1.3 satellite embedding model with open weights and precomputed global 10m embeddings plus a browser-based Explorer
Ai2 released the OlmoEarth v1.3 Base model (weights, pretraining data and code under the OlmoEarth Artifact License), precomputed annual global embeddings at 10m resolution (128- and 64-dimension, hosted on Source Cooperative; 2023-2025 global, 2017-2022 select countries, roughly 150 TB per year), a no-sign-in OlmoEarth Embeddings Explorer web tool, a technical report, and the new PLANTEUR tropical crop-type benchmark. Ai2 reports its 128-dim embeddings score highest on 20 of 27 evals against AlphaEarth Foundations and Tessera v2 (64-dim leads on 18 of 27). It also reports v1.3 trained with 53% of the compute of v1.2 and 18% of v1.
Entities: Ai2 (Allen Institute for AI), OlmoEarth v1.3, OlmoEarth Embeddings Explorer, OlmoEarth Studio, AlphaEarth Foundations, Tessera v2
1 primary
What happened
Ai2 released OlmoEarth v1.3 Base, with weights, pretraining data and code under its own OlmoEarth Artifact License. It also published precomputed annual global embeddings at 10m resolution (128 and 64 dimensions, hosted on Source Cooperative, roughly 150 TB per year). Coverage is global for 2023 to 2025 and select countries for 2017 to 2022. Alongside these came a no-sign-in browser Explorer, a technical report and a new tropical crop-type benchmark called PLANTEUR. Ai2 says its 128-dim embeddings lead on 20 of 27 evaluations against AlphaEarth Foundations and Tessera v2, and that v1.3 used 53% of the compute of v1.2.
Why it matters
Researchers, NGOs and small developers who map crops, land cover or illegal mining can now use ready-made global embeddings without training or running a large model themselves. That lowers the cost of entry and gives an open option next to Google's AlphaEarth. The practical limits are data size (150 TB per year is heavy to download or store) and the licence, which is custom rather than a standard open-source one, so commercial users should read it. Whether the quality holds on real, local tasks is not yet shown.
What is noise
"State-of-the-art" rests on Ai2's own benchmark selection, and the compute savings are also self-reported. A win on 20 of 27 evaluations also means it lost or tied on 7, and the 64-dim version leads on only 18. "Fully open" is a stretch for a bespoke licence, and no independent comparison yet shows how large the gaps are or whether they matter in practice.
Watch next
- 01Independent benchmarks or replications comparing OlmoEarth, AlphaEarth and Tessera v2 on tasks and regions outside Ai2's own evaluation suite, especially PLANTEUR.
- 02Real adoption signals: Hugging Face and Source Cooperative download counts, Explorer usage, and named conservation, agriculture or government projects using the embeddings within the next few months.
- 03Scrutiny of the OlmoEarth Artifact License terms for commercial use, plus Ai2's plans to extend 2017 to 2022 coverage beyond select countries and any response from Google on AlphaEarth access.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680
- OpenAI launches GPT-6 Sol and Luna at half the token price of GPT-5.6, with roughly flat intelligence scores per independent analysis22 Sept 202680