Ai2 Releases Open Satellite Embeddings for Every 10m Pixel on Earth

The Allen Institute for AI (Ai2) has released OlmoEarth v1.3, a model that boils a year of satellite imagery into a list of 128 numbers for every 10-metre square of land on Earth. Ai2 announced the release on 22 October together with precomputed maps of those numbers for the whole planet for 2023, 2024 and 2025, and a free browser tool to explore them with “no sign-in, and no advanced technical knowledge required”.

Ai2 says its embeddings beat two rival satellite embedding products on 20 of 27 evaluations. That figure comes from Ai2’s own testing.

The weights, pretraining data and pretraining code are all published, according to Ai2, and the embeddings are hosted for download on Source Cooperative.

Each pixel gets a 128-number summary of a year of satellite data

An embedding is a compact numerical description. OlmoEarth v1.3 “reads a year of radar (Sentinel-1) and multispectral (Sentinel-2 and Landsat) satellite time series and produces one 128-dimensional vector for every 10-meter pixel on Earth”, Ai2 says. The first 64 numbers also work on their own as a smaller version.

The point is to remove the expensive step. Until now, Ai2 says, the main way to use OlmoEarth was to fine-tune a model for each task in its OlmoEarth Studio platform, which “requires GPU access and the ability to download and store the imagery”. With precomputed embeddings, a crop map or land-cover map “becomes a lightweight modeling problem: train a small classifier on top and run it locally”.

The model card describes a 128-million-parameter vision transformer, with a smaller component that compresses its output to 128 dimensions. Ai2 gives four example uses: “classifying smallholder crops in Kenya, detecting brick kilns west of Dhaka, assessing fire damage in Indonesia, and tracking gold mining in the Peruvian Amazon”.

The archive holds roughly 150 TB per year, covering land and nearby coastal waters. The 2023 to 2025 layers are global; 2017 to 2022 cover only selected countries, Ai2 says. Each 20 km tile comes with provenance tables recording which satellite scenes went into it.

The free explorer runs in a browser with no account

The OlmoEarth Embeddings Explorer renders the embeddings as a false-colour map of the planet. According to Ai2, users can flip between years, pull up the underlying Sentinel-2 imagery, click a spot to find similar-looking places, or drop a few labelled points to train a classifier and export the result as a GeoTIFF file. “No account, no setup required,” the company says, and a first pass “takes seconds” with “no GPU needed”.

Ai2’s showcase is the Peruvian Amazon. Illegal gold mining around La Pampa in Madre de Dios peaked in 2017 and 2018, clearing more than 1,200 hectares of forest in the buffer zone of the Tambopata National Reserve, Ai2 says, citing Amazon Conservation’s MAAP monitoring programme. Peru’s Operation Mercurio shut most of it down in February 2019.

Comparing each pixel’s embedding across years, with no model training, shows the effect, Ai2 says. “The crackdown is visible in the embeddings,” Ai2 writes. But pockets of activity continued, and Ai2 notes that, according to the Associated Press, Peru sent navy patrols back into the reserve in August 2026 after prosecutors found mining inside Tambopata itself.

Ai2 says v1.3 cost about half as much to train as v1.2

Ai2 says OlmoEarth v1.3 took 53% of the computing time of v1.2 and 18% of the original v1, with 50% and 22% of their carbon emissions, while matching v1.2’s performance when fine-tuned. The final pre-training run “would cost roughly $2,720 of rented H100 time”, the company estimates.

The training code is on GitHub. Ai2’s technical report says lower costs are meant to let more researchers work on geospatial foundation models.

The release fits a wider push to put Earth data in more hands, from Google’s experimental Earth AI geospatial reasoning agent to Vaisala’s use of NVIDIA’s satellite-fed lightning nowcasting. More projects like it are collected in TSN’s Tech for Good hub.

Ai2’s benchmark lead is real on its own tests but narrower than “20 of 27” suggests

Ai2 tested its embeddings against Google DeepMind’s AlphaEarth Foundations, which it retrieved through Google Earth Engine, and Tessera v2, across 27 evaluations on six continents. Ai2 says its 128-dimensional embeddings “score highest on 20 of 27 evals and rank highest overall”.

The technical report shows a more mixed picture. The comparison is not size-matched: AlphaEarth’s embeddings have 64 dimensions, half the size of Ai2’s main product. Ai2’s own 64-dimensional version still leads on 18 of 27, it says. On one crop-mapping benchmark, PASTIS, the 128-dimensional version beats Tessera v2 by 58.6 to 58.3, a narrow margin. On Ethiopian crop data, the report says, “TESSERA v2 consistently outperforms OlmoEarth v1.3”. One of the evaluations, PLANTEUR, is a tropical crop benchmark Ai2 introduced alongside the release.

The report also says the embeddings it tested are “not bit-identical” to the ones released, because the public files were produced with a different inference method; Ai2 measured a negligible gap on PASTIS. None of this suggests the results are wrong, but no outside group has yet repeated them.

Ai2’s “open” licence bars military and extractive-industry use

Ai2 calls the release “fully open”, and the weights, data and code are published. The licence carries conditions, however. The OlmoEarth Artifact License, Ai2’s technical report says, “restricts use for military, defense-related, and extractive industry applications”.

That puts it short of an unrestricted open-source licence, even though anyone can download and inspect the model. For conservation groups and agricultural researchers, the restriction is unlikely to bite; the Peru example is monitoring mining, not doing it. Mining companies and defence users, by contrast, would need to read the terms closely before building on it.

The more immediate test is the map itself. Ai2 has not said when a 2026 layer will follow. When it does, the same year-on-year comparison will show whether the navy patrols Peru sent back into Tambopata in August pushed the miners out.

Sources

• Ai2: OlmoEarth v1.3: State-of-the-art embeddings for the whole planet, 22 October 2026
• Hugging Face blog (Ai2): OlmoEarth v1.3: State-of-the-art embeddings for the whole planet, 22 October 2026
• Ai2: OlmoEarth v1.3 technical report (PDF)
• Hugging Face: allenai/OlmoEarth-v1_3-Base model card
• GitHub: allenai/olmoearth_pretrain
• Source Cooperative: OlmoEarth v1.3 Embeddings
• MAAP: La Pampa and Operation Mercurio

Related stories

• Google Expands Its Experimental Earth AI Geospatial Reasoning Agent
• Vaisala Says NVIDIA AI Stretches Lightning Warnings to Three Hours
• Google DeepMind Gives Up to 200 European Researchers Access to Its Frontier Models, Plus $3M in Compute Credits