The product is the precomputation, not the model

OlmoEarth v1.3 is, on paper, a modest model: a 128-million-parameter vision transformer, smaller than the systems that usually get launch coverage. What Ai2 actually shipped is the work around it. A year of Sentinel-1 radar, Sentinel-2 and Landsat observations for every 10-metre patch of land in the archive has already been run through the model — about 1.3 trillion embeddings for a single year, by Ai2's count — and the resulting vectors sit in an open Zarr store on Source Cooperative that any Zarr client can read directly from object storage, with no account, API key or tile server. The Explorer on top of it needs no sign-in: pick a place, compare years, find similar places, label a handful of points to train a small classifier, export a GeoTIFF.

That changes who can start. The precomputed route is new: before this release, embeddings were something you computed for your own area of interest — OlmoEarth Studio has offered custom embedding exports since April 2026, and the open models can be run locally — while earlier OlmoEarth generations were deployed principally by fine-tuning inside Studio, which meant GPU access and storing the imagery yourself. The archive route skips both for a first map. No API key does not mean no requirements, though: the store is built for area reads, its smallest transfer unit is a compressed chunk of roughly 3 MB, and whatever you pull still has to be stored and processed on your side.

Ai2's launch example is illegal gold mining around La Pampa in Peru's Madre de Dios region: comparing precomputed years showed the earlier spread of mining and continued change in the old zone, with no task-specific training at all. As a demonstration of the format, it is well chosen. It is also Ai2's own example, presented by Ai2, and it shows change detection working, not a measurement independently checked against ground survey.

Read the scoreboard in two halves

The headline count is Ai2's: in its announcement, the 128-dimensional embeddings scored highest on 20 of 27 evaluations against two named rivals, Google's AlphaEarth Foundations and TESSERA v2, with the smaller 64-dimensional version leading on 18. The technical report's own summary is narrower and checkable. Using a single kNN-20 classifier, it puts the 128-dimensional embeddings best on 10 of 11 evaluations. Under the AEF evaluation protocol the report follows, they beat both rivals on seven of eight datasets; the report adds that the 128-dimensional embeddings stay best on seven of eight under every linear classifier it tried, and on six of eight under random forests and an MLP. The exception is consistent: on the Ethiopia crops dataset, TESSERA v2 wins. The evaluations are Ai2's and the report is Ai2's. A test suite the maker ran, on which the maker's model leads, is evidence. It is not independent evidence.

The win count also hides the margins. On PASTIS, a standard crop-mapping dataset, Ai2 reports 58.6 mIoU for its 128-dimensional embeddings against 58.3 for TESSERA v2 — a 0.3-point difference. Counting both narrow and wide leads as one win is useful for a scorecard, but it is not a measure of how large the advantage is. One further caveat sits in the report's methods: the evaluated embeddings were computed by tiling the vision transformer, not with the Searchlight pipeline used for the released archive; Ai2 measured the difference on PASTIS and calls it negligible.

PLANTEUR, the new benchmark Ai2 introduced with this release, is where the margin is widest. It is a crop-type dataset built from labels in five French overseas territories — Martinique, Réunion, French Guiana, Guadeloupe and Mayotte — and its held-out-territory test trains a classifier on four territories and tests it on the fifth. The report calls the design "leave-one-island-out", although one of the five, French Guiana, sits on the South American mainland. On that test v1.3 scored 24.9 mIoU against 19.1 for TESSERA v2 and 15.8 for AlphaEarth Foundations. That is a real margin on a sensible generalisation test. It is also a dataset and evaluation Ai2 designed, so the result is not an independent demonstration of general performance across countries or land-cover types.

The second half of the scoreboard points the other way. Rank every evaluated model across all of the report's fine-tuning tasks — a mix of research benchmarks and partner tasks, omitting task-model combinations a model did not run — and v1.3, without its Attention Pooler, has a mean rank of 4.34. The previous version, v1.2 Base, ranks at 4.02; v1 Base ranks at 4.74. The split matters: v1.2 is clearly better on the research benchmarks (mean rank 2.60 against v1.3's 4.50), while v1.3 is better on the partner tasks (4.23 against 4.97). Ai2's own description is that v1.3 maintains v1.2's fine-tuning quality, which is close to fair. But a reader deciding whether to rebuild an existing fine-tuning pipeline on v1.3 gets no clear improvement from the report's own numbers. The upgrade is in the embeddings and the archive, not in the model you train yourself.

Cheap to train, expensive to run over a planet

Ai2 calls v1.3 its cheapest OlmoEarth to train, and in GPU-hours the claim is plain: the final pretraining run for the Base model took 539 H100 GPU hours, 53% of v1.2's figure and 18% of v1's. Ai2 estimates that reproducing that final run would cost roughly $2,720 in rented H100 time. That figure will travel, and it needs its limits attached: it prices one final training run, not the experiments, data preparation, storage or serving around it.

The figure beside it matters more for understanding the release. Computing one year of global embeddings is estimated at about 14,000 H100 GPU hours — roughly 26 times the GPU-hours of that final pretraining run, and itself an estimate, since Ai2's inference fleet mixes H100 and A100 machines. Each year of archive at global coverage is about 150 TB, and Ai2 puts the full 2017-to-2025 set at well over a petabyte.

So the economics are a transfer, not a disappearance. Ai2 paid the planetary inference bill once so that thousands of smaller teams do not each pay a local version of it. That is a genuinely useful thing for a nonprofit institute to do, and it is also a dependency: the map you get for free reflects Ai2's model, Ai2's year boundaries and Ai2's processing choices, including the Searchlight inference Ai2 says removed tiling artefacts from the released embeddings. If your question needs a specific season, a different time window or a resolution the archive does not carry, you are back to computing embeddings yourself, in Studio or on your own hardware — where the fine-tuning comparison with v1.2 is, on Ai2's own numbers, a draw.

"Whole planet", with the map edges drawn in

The coverage limits are stated in Ai2's own materials and are worth keeping attached to the headline. Two years, 2024 and 2025, are global. The third, 2023, is near-global, with the remaining regions still being filled. The six earlier years, 2017 to 2022, cover selected countries, with backfilling in progress. The covered area runs from 59.5 degrees south to 83.6 degrees north and excludes Antarctica, and a narrow strip at the antimeridian has no embeddings at all. For most land-use questions none of that binds. For anyone working on polar change, or on a country outside the early-years list, the precomputed planet is not yet the planet.

The licence draws a second boundary, and its text is more specific than the summaries suggest. Section 2 of the OlmoEarth Artifact License — the licence that ships with the dataset and covers the model weights, dataset and code — bars any use of the artifacts, or of derivatives made from them, for military and defence-related applications, including weapons development, military operations, intelligence gathering, and human surveillance and policing. It also bars extractive activities involving the removal of raw materials from the earth, including planning or facilitating the extraction of oil, natural gas and minerals through drilling, mining and deforestation. Section 3 requires anyone distributing derivatives to carry those same restrictions downstream.

Read precisely, the licence restricts uses, not whole industries. Ai2's own demonstration — detecting illegal gold mining so it can be policed — is an environmental-monitoring use, not the planning or facilitation of extraction, and nothing in the licence text makes it a breach. The real position is narrower and more durable than either slogan: this is a use-restricted licence, not unrestricted open source, and whether a given project is permitted is a question its exact application, not its sector, answers.

Sources

  1. [1] Ai2 (Allen Institute for AI), “OlmoEarth v1.3 Embeddings”, technical report (allenai.org, 22 October 2026) — read in full for this article: the model and training (§2); the embedding evaluations (§3.1, Tables 1–2: best on seven of eight AEF datasets, the Ethiopia exception, PASTIS 58.6 mIoU against TESSERA v2's 58.3, PLANTEUR held-out-territory scores of 24.9 against 19.1 and 15.8); the fine-tuning ranking (§3.2.2: mean ranks 4.34, 4.02 and 4.74, with the research/partner splits); deployment and coverage (§4, including Searchlight in §4.2, distribution in §4.7 — about 150 TB per year, well over a petabyte for 2017–2025, no API key or tile server — and the Explorer in §4.8); and the environmental figures (§5: 539 H100 GPU hours for the final pretraining run, Table 5; about 14,000 H100 hours per year of global embeddings, §5.2).Read source
  2. [2] Ai2, OlmoEarth-v1_3-Base model card (Hugging Face) — a ViT-Base foundation model with 128 million parameters for remote sensing over Sentinel-1, Sentinel-2 and Landsat images or image time series; a Perceiver of 14.3 million parameters; a linear student of 0.1 million parameters projecting 768-dimensional tokens to 128; the statement that this model, student included, generated OlmoEarth's global annualized embeddings; licensed under the OlmoEarth Artifact License and intended for use in accordance with Ai2's Responsible Use Guidelines.Read source
  3. [3] Ai2, OlmoEarth v1.3 Embeddings dataset (Source Cooperative) — the release-day coverage note (2024 and 2025 global; 2023 near-global, with the remaining regions being filled; 2017–2022 cover selected countries and are being backfilled), the coverage extent of 59.5°S to 83.6°N excluding Antarctica, public anonymous access through the S3-compatible data proxy with no account or API key, the Zarr v3 layout and the dataset caveats, including the antimeridian strip with no embeddings.Read source
  4. [4] OlmoEarth Artifact License, full text as shipped with the dataset (Source Cooperative) — §1 use rights; §2 use restrictions: no use of the artifacts or derivatives for military and defence-related applications, including weapons development, military operations, intelligence gathering, or human surveillance and policing, and none for extractive activities involving the removal of raw materials from the earth, including planning or facilitating the extraction of oil, natural gas and minerals through drilling, mining and deforestation; §3 distribution, including citing Ai2 and carrying the §2 restrictions into all downstream distributions of derivatives; §7 other agreements.Read source
  5. [5] Ai2, “Another way to view our planet” (narrative.allen.ai, 22 October 2026) — Ai2's own description of the embeddings (about 1.3 trillion for the entire planet for one year; every 10-metre pixel of most of the world's land described as an embedding) and of the Explorer's free, no-sign-in access.Read source
  6. [6] Ai2, “Introducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis” (allenai.org blog, 23 April 2026) — since April 2026, OlmoEarth Studio has let users compute and export embedding vectors for their own area of interest, time range, encoder variant, resolution and imagery sources, computed on demand rather than pulled from a precomputed global archive.Read source
  7. [7] Unite.AI, “Ai2 Releases OlmoEarth V1.3 With Global 10M Satellite Embeddings” (22 October 2026) — account of Ai2's announcement, used only for the announcement's own figures and descriptions that are not tabulated in the technical report: the 20-of-27 and 18-of-27 headline counts, Ai2's roughly $2,720 estimate for reproducing the final pretraining run in rented H100 time, the Explorer's classifier training from a handful of labelled points and GeoTIFF export, and the La Pampa, Peru demonstration details.Read source